Datasets:

saier

/

unarXive_citrec

like 7

4ddf79b3-fcfc-4730-a50a-8a67c1cbeeb9

We validate our approach, MODEST (Mobile Object Detection with Ephemerality and Self-Training) on the Lyft Level 5 Perception Dataset [1]} and the nuScenes Dataset [2]} with various types of detectors [3]}, [4]}, [5]}, [6]}. We demonstrate that MODEST yields remarkably accurate mobile object detectors, comparable to their supervised counterparts. Concretely, our contributions are three-fold:

[2]

https://openalex.org/W3035574168

58439ca8-5b5b-4830-ab80-cc8b7bff123a

We validate our approach, MODEST (Mobile Object Detection with Ephemerality and Self-Training) on the Lyft Level 5 Perception Dataset [1]} and the nuScenes Dataset [2]} with various types of detectors [3]}, [4]}, [5]}, [6]}. We demonstrate that MODEST yields remarkably accurate mobile object detectors, comparable to their supervised counterparts. Concretely, our contributions are three-fold:

[3]

https://openalex.org/W2949708697

7bdf97ba-6df9-4dd2-9558-46beedefdcfb

We validate our approach, MODEST (Mobile Object Detection with Ephemerality and Self-Training) on the Lyft Level 5 Perception Dataset [1]} and the nuScenes Dataset [2]} with various types of detectors [3]}, [4]}, [5]}, [6]}. We demonstrate that MODEST yields remarkably accurate mobile object detectors, comparable to their supervised counterparts. Concretely, our contributions are three-fold:

[5]

https://openalex.org/W2963727135

15729042-80fe-44d2-99bf-5bf143654f06

We validate our approach, MODEST (Mobile Object Detection with Ephemerality and Self-Training) on the Lyft Level 5 Perception Dataset [1]} and the nuScenes Dataset [2]} with various types of detectors [3]}, [4]}, [5]}, [6]}. We demonstrate that MODEST yields remarkably accurate mobile object detectors, comparable to their supervised counterparts. Concretely, our contributions are three-fold:

[6]

https://openalex.org/W2897529137

deb1491b-509e-4a06-b024-3412e48093e2

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[1]

https://openalex.org/W2560609797

290c4146-dd84-43dc-8b20-d18ebafc466c

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[2]

https://openalex.org/W2963121255

afa5aa97-7c96-4c38-9671-72867b97113a

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[3]

https://openalex.org/W2964062501

2dd7949d-2f09-4b0f-87b3-ddad673b4318

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[4]

https://openalex.org/W2949708697

af598327-b522-474c-81f1-8abca08c205c

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[6]

https://openalex.org/W2963727135

43a2ce96-2d3d-424e-9fdd-1a64b1e4a2ac

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[7]

https://openalex.org/W3031752193

8d15e1e7-ae65-446b-9b4b-ad7a5b0b8d20

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[8]

https://openalex.org/W2798965597

b05f478d-cb02-4ff9-9b72-9c20119c2bb4

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[9]

https://openalex.org/W2897529137

be69b078-04db-4920-a697-3f68ba5a86ca

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[10]

https://openalex.org/W3034314779

7cf0666a-0826-44e0-8de5-91a7f0a0e979

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[11]

https://openalex.org/W3108486966

c017a118-822a-4a63-bd9d-edfa22da1905

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[13]

https://openalex.org/W2555618208

b68be34f-ec6b-46be-9903-ff17d45aea39

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[14]

https://openalex.org/W2150066425

1597a5eb-3b90-42fe-bbb8-57468ae2aec0

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[15]

https://openalex.org/W2115579991

cde76ff3-1388-438a-916e-f8e36d00acbb

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[17]

https://openalex.org/W3035574168

378886c3-6ecc-4622-abc9-8871bb18e481

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[18]

https://openalex.org/W2955189650

aa0a60cf-ecae-4d98-a70b-77a922e9d22c

3D object detection and existing datasets. Most existing 3D object detectors take 3D point clouds generated by LiDAR as input. They either consist of specialized neural architectures that can operate on point clouds directly [1]}, [2]}, [3]}, [4]}, [5]} or voxelize the point clouds to leverage 2D or 3D convolutional neural architectures [6]}, [7]}, [8]}, [9]}, [10]}, [11]}, [12]}, [13]}. Regardless of the architectures, they are trained using supervision and their performances hinges directly on the training dataset. However, the limited variety of objects and driving conditions in existing autonomous driving datasets [14]}, [15]}, [16]}, [17]}, [18]} impedes the generalizability of the resulting detectors [19]}.

[19]

https://openalex.org/W3034975685

e5e70a45-16d6-4838-8af0-4369d30eac00

Unsupervised Object Discovery in 2D/3D. Our work follows prior work on discovering objects both from 2D images as well as from 3D data. A first step in object discovery is to identify candidate objects, or “proposals” from a single scene/image. For 2D images, this is typically done by segmenting the image using appearance cues[1]}, [2]}, [3]}, [4]}, but color variations and perspective effects make this difficult. Tian  [5]} exploits the correspondence between images and 3D point clouds to detect objects in 2D. In 3D scenes, one can use 3D information such as surface normals [6]}, [7]}, [2]}, [9]}, [10]}, [11]}, [12]}, [13]}. One can also use temporal changes such as motion [14]}, [15]}, [16]}, [1]}, [9]}. Our work combines effective 3D information with cues from changes in the scene over time to detect mobile objects[10]}, [11]}, [21]}. In particular, similar to our approach, Herbst et al.[10]}, [11]} reconstruct the same scene at various times and carve out dissimilar regions as mobile objects. We use the analogous idea of ephemerality as proposed by Barnes et al.[24]}. We show in our work that this idea yields a surprisingly accurate set of initial objects. In addition, we also leverage other common-sense rules such as locations of the objects (e.g. objects should stay on the ground) [25]}, [26]}, [27]}, [28]} or shapes of an object (e.g. objects should be compact) [12]}, [27]}, [28]}. However, crucially, we do not just stop at this proposal stage. Instead, we use these seed labels to train an object detector through multiple rounds of self-training. This effectively identifies objects consistent across multiple scenes. While previous work has attempted to use this consistency cue[3]}, [4]}, [34]}, [6]}, [7]}, [37]}, [38]} (including co-segmentation[39]}, [40]}, [41]}), prior work typically uses clustering to accomplish this. In contrast, we demonstrate that neural network training and self-training provides a very strong signal and substantially improves the quality of the proposals or seed labels.

[3]

https://openalex.org/W2979654309

6f935b40-ccfc-4b4a-82b7-c3ffbbd93fad

Unsupervised Object Discovery in 2D/3D. Our work follows prior work on discovering objects both from 2D images as well as from 3D data. A first step in object discovery is to identify candidate objects, or “proposals” from a single scene/image. For 2D images, this is typically done by segmenting the image using appearance cues[1]}, [2]}, [3]}, [4]}, but color variations and perspective effects make this difficult. Tian  [5]} exploits the correspondence between images and 3D point clouds to detect objects in 2D. In 3D scenes, one can use 3D information such as surface normals [6]}, [7]}, [2]}, [9]}, [10]}, [11]}, [12]}, [13]}. One can also use temporal changes such as motion [14]}, [15]}, [16]}, [1]}, [9]}. Our work combines effective 3D information with cues from changes in the scene over time to detect mobile objects[10]}, [11]}, [21]}. In particular, similar to our approach, Herbst et al.[10]}, [11]} reconstruct the same scene at various times and carve out dissimilar regions as mobile objects. We use the analogous idea of ephemerality as proposed by Barnes et al.[24]}. We show in our work that this idea yields a surprisingly accurate set of initial objects. In addition, we also leverage other common-sense rules such as locations of the objects (e.g. objects should stay on the ground) [25]}, [26]}, [27]}, [28]} or shapes of an object (e.g. objects should be compact) [12]}, [27]}, [28]}. However, crucially, we do not just stop at this proposal stage. Instead, we use these seed labels to train an object detector through multiple rounds of self-training. This effectively identifies objects consistent across multiple scenes. While previous work has attempted to use this consistency cue[3]}, [4]}, [34]}, [6]}, [7]}, [37]}, [38]} (including co-segmentation[39]}, [40]}, [41]}), prior work typically uses clustering to accomplish this. In contrast, we demonstrate that neural network training and self-training provides a very strong signal and substantially improves the quality of the proposals or seed labels.

[4]

https://openalex.org/W1919709169

48938c3e-838c-4ff0-9732-d22435682279

Unsupervised Object Discovery in 2D/3D. Our work follows prior work on discovering objects both from 2D images as well as from 3D data. A first step in object discovery is to identify candidate objects, or “proposals” from a single scene/image. For 2D images, this is typically done by segmenting the image using appearance cues[1]}, [2]}, [3]}, [4]}, but color variations and perspective effects make this difficult. Tian  [5]} exploits the correspondence between images and 3D point clouds to detect objects in 2D. In 3D scenes, one can use 3D information such as surface normals [6]}, [7]}, [2]}, [9]}, [10]}, [11]}, [12]}, [13]}. One can also use temporal changes such as motion [14]}, [15]}, [16]}, [1]}, [9]}. Our work combines effective 3D information with cues from changes in the scene over time to detect mobile objects[10]}, [11]}, [21]}. In particular, similar to our approach, Herbst et al.[10]}, [11]} reconstruct the same scene at various times and carve out dissimilar regions as mobile objects. We use the analogous idea of ephemerality as proposed by Barnes et al.[24]}. We show in our work that this idea yields a surprisingly accurate set of initial objects. In addition, we also leverage other common-sense rules such as locations of the objects (e.g. objects should stay on the ground) [25]}, [26]}, [27]}, [28]} or shapes of an object (e.g. objects should be compact) [12]}, [27]}, [28]}. However, crucially, we do not just stop at this proposal stage. Instead, we use these seed labels to train an object detector through multiple rounds of self-training. This effectively identifies objects consistent across multiple scenes. While previous work has attempted to use this consistency cue[3]}, [4]}, [34]}, [6]}, [7]}, [37]}, [38]} (including co-segmentation[39]}, [40]}, [41]}), prior work typically uses clustering to accomplish this. In contrast, we demonstrate that neural network training and self-training provides a very strong signal and substantially improves the quality of the proposals or seed labels.

[12]

https://openalex.org/W2049776679

5e2cd74a-bfe3-43f7-8752-edaba6f5baae

Unsupervised Object Discovery in 2D/3D. Our work follows prior work on discovering objects both from 2D images as well as from 3D data. A first step in object discovery is to identify candidate objects, or “proposals” from a single scene/image. For 2D images, this is typically done by segmenting the image using appearance cues[1]}, [2]}, [3]}, [4]}, but color variations and perspective effects make this difficult. Tian  [5]} exploits the correspondence between images and 3D point clouds to detect objects in 2D. In 3D scenes, one can use 3D information such as surface normals [6]}, [7]}, [2]}, [9]}, [10]}, [11]}, [12]}, [13]}. One can also use temporal changes such as motion [14]}, [15]}, [16]}, [1]}, [9]}. Our work combines effective 3D information with cues from changes in the scene over time to detect mobile objects[10]}, [11]}, [21]}. In particular, similar to our approach, Herbst et al.[10]}, [11]} reconstruct the same scene at various times and carve out dissimilar regions as mobile objects. We use the analogous idea of ephemerality as proposed by Barnes et al.[24]}. We show in our work that this idea yields a surprisingly accurate set of initial objects. In addition, we also leverage other common-sense rules such as locations of the objects (e.g. objects should stay on the ground) [25]}, [26]}, [27]}, [28]} or shapes of an object (e.g. objects should be compact) [12]}, [27]}, [28]}. However, crucially, we do not just stop at this proposal stage. Instead, we use these seed labels to train an object detector through multiple rounds of self-training. This effectively identifies objects consistent across multiple scenes. While previous work has attempted to use this consistency cue[3]}, [4]}, [34]}, [6]}, [7]}, [37]}, [38]} (including co-segmentation[39]}, [40]}, [41]}), prior work typically uses clustering to accomplish this. In contrast, we demonstrate that neural network training and self-training provides a very strong signal and substantially improves the quality of the proposals or seed labels.

[24]

https://openalex.org/W2963170338

90123595-9cfd-4181-bc3a-885955566240

Unsupervised Object Discovery in 2D/3D. Our work follows prior work on discovering objects both from 2D images as well as from 3D data. A first step in object discovery is to identify candidate objects, or “proposals” from a single scene/image. For 2D images, this is typically done by segmenting the image using appearance cues[1]}, [2]}, [3]}, [4]}, but color variations and perspective effects make this difficult. Tian  [5]} exploits the correspondence between images and 3D point clouds to detect objects in 2D. In 3D scenes, one can use 3D information such as surface normals [6]}, [7]}, [2]}, [9]}, [10]}, [11]}, [12]}, [13]}. One can also use temporal changes such as motion [14]}, [15]}, [16]}, [1]}, [9]}. Our work combines effective 3D information with cues from changes in the scene over time to detect mobile objects[10]}, [11]}, [21]}. In particular, similar to our approach, Herbst et al.[10]}, [11]} reconstruct the same scene at various times and carve out dissimilar regions as mobile objects. We use the analogous idea of ephemerality as proposed by Barnes et al.[24]}. We show in our work that this idea yields a surprisingly accurate set of initial objects. In addition, we also leverage other common-sense rules such as locations of the objects (e.g. objects should stay on the ground) [25]}, [26]}, [27]}, [28]} or shapes of an object (e.g. objects should be compact) [12]}, [27]}, [28]}. However, crucially, we do not just stop at this proposal stage. Instead, we use these seed labels to train an object detector through multiple rounds of self-training. This effectively identifies objects consistent across multiple scenes. While previous work has attempted to use this consistency cue[3]}, [4]}, [34]}, [6]}, [7]}, [37]}, [38]} (including co-segmentation[39]}, [40]}, [41]}), prior work typically uses clustering to accomplish this. In contrast, we demonstrate that neural network training and self-training provides a very strong signal and substantially improves the quality of the proposals or seed labels.

[34]

https://openalex.org/W3106670560

1805295a-986f-42da-835a-05f1dc515ac2

Unsupervised Object Discovery in 2D/3D. Our work follows prior work on discovering objects both from 2D images as well as from 3D data. A first step in object discovery is to identify candidate objects, or “proposals” from a single scene/image. For 2D images, this is typically done by segmenting the image using appearance cues[1]}, [2]}, [3]}, [4]}, but color variations and perspective effects make this difficult. Tian  [5]} exploits the correspondence between images and 3D point clouds to detect objects in 2D. In 3D scenes, one can use 3D information such as surface normals [6]}, [7]}, [2]}, [9]}, [10]}, [11]}, [12]}, [13]}. One can also use temporal changes such as motion [14]}, [15]}, [16]}, [1]}, [9]}. Our work combines effective 3D information with cues from changes in the scene over time to detect mobile objects[10]}, [11]}, [21]}. In particular, similar to our approach, Herbst et al.[10]}, [11]} reconstruct the same scene at various times and carve out dissimilar regions as mobile objects. We use the analogous idea of ephemerality as proposed by Barnes et al.[24]}. We show in our work that this idea yields a surprisingly accurate set of initial objects. In addition, we also leverage other common-sense rules such as locations of the objects (e.g. objects should stay on the ground) [25]}, [26]}, [27]}, [28]} or shapes of an object (e.g. objects should be compact) [12]}, [27]}, [28]}. However, crucially, we do not just stop at this proposal stage. Instead, we use these seed labels to train an object detector through multiple rounds of self-training. This effectively identifies objects consistent across multiple scenes. While previous work has attempted to use this consistency cue[3]}, [4]}, [34]}, [6]}, [7]}, [37]}, [38]} (including co-segmentation[39]}, [40]}, [41]}), prior work typically uses clustering to accomplish this. In contrast, we demonstrate that neural network training and self-training provides a very strong signal and substantially improves the quality of the proposals or seed labels.

[39]

https://openalex.org/W1984034752

edbaf067-105d-4b58-b8b3-1e476b4590cf

Unsupervised Object Discovery in 2D/3D. Our work follows prior work on discovering objects both from 2D images as well as from 3D data. A first step in object discovery is to identify candidate objects, or “proposals” from a single scene/image. For 2D images, this is typically done by segmenting the image using appearance cues[1]}, [2]}, [3]}, [4]}, but color variations and perspective effects make this difficult. Tian  [5]} exploits the correspondence between images and 3D point clouds to detect objects in 2D. In 3D scenes, one can use 3D information such as surface normals [6]}, [7]}, [2]}, [9]}, [10]}, [11]}, [12]}, [13]}. One can also use temporal changes such as motion [14]}, [15]}, [16]}, [1]}, [9]}. Our work combines effective 3D information with cues from changes in the scene over time to detect mobile objects[10]}, [11]}, [21]}. In particular, similar to our approach, Herbst et al.[10]}, [11]} reconstruct the same scene at various times and carve out dissimilar regions as mobile objects. We use the analogous idea of ephemerality as proposed by Barnes et al.[24]}. We show in our work that this idea yields a surprisingly accurate set of initial objects. In addition, we also leverage other common-sense rules such as locations of the objects (e.g. objects should stay on the ground) [25]}, [26]}, [27]}, [28]} or shapes of an object (e.g. objects should be compact) [12]}, [27]}, [28]}. However, crucially, we do not just stop at this proposal stage. Instead, we use these seed labels to train an object detector through multiple rounds of self-training. This effectively identifies objects consistent across multiple scenes. While previous work has attempted to use this consistency cue[3]}, [4]}, [34]}, [6]}, [7]}, [37]}, [38]} (including co-segmentation[39]}, [40]}, [41]}), prior work typically uses clustering to accomplish this. In contrast, we demonstrate that neural network training and self-training provides a very strong signal and substantially improves the quality of the proposals or seed labels.

[41]

https://openalex.org/W2964283970

97a869b7-eea0-4483-af9c-f4943fc47445

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[2]

https://openalex.org/W3035160371

406925b9-7399-474a-9e8d-c7206f2b0f2b

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[3]

https://openalex.org/W2895281799

f5680c4a-8d2f-44f0-85da-be38801d80a2

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[4]

https://openalex.org/W2963240485

644401b9-8347-4976-a4be-dcf2227fb4e9

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[5]

https://openalex.org/W2985406498

3459c54f-2461-48c6-b1db-7f8a040fcb01

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[6]

https://openalex.org/W2970092410

462932a8-1591-4245-93ff-b3ad459c827c

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[7]

https://openalex.org/W3108566666

c2568917-4783-4a55-8b3e-3b87a40a90d2

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[8]

https://openalex.org/W3175269419

325c9b93-4b28-44fb-a349-27fddfc17e6f

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[9]

https://openalex.org/W3108975329

45224ba8-16cf-483d-9402-ac524cc1affd

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[10]

https://openalex.org/W3128167848

18a55fb5-66f8-4a48-9454-336be5b97baf

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[12]

https://openalex.org/W3204397973

2187df7f-b83a-453d-82e9-e7ecc28e2ed4

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[13]

https://openalex.org/W2963735582

aab9c985-e4c9-47b2-bc42-d971870f9650

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[14]

https://openalex.org/W3004146535

919489a8-2de9-410e-a23b-bf56f62351b0

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[15]

https://openalex.org/W2963096987

08b2a223-ae79-48b5-8d07-f483023db56d

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[16]

https://openalex.org/W2575671312

42d9a250-788a-4d70-a460-142456129ac2

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[17]

https://openalex.org/W2145494108

4890480e-9f78-4759-adcc-23994bc86779

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[18]

https://openalex.org/W2978426779

75c1a013-a367-4f90-9f02-ed69743c8bc8

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[19]

https://openalex.org/W2989700832

6fc51ad7-ebf6-4dfc-880b-e90b52febbc9

Self-training, semi-supervised and self-supervised learning. When training our detector, we use self-training, which has been shown to be highly effective for semi-supervised learning[1]}, [2]}, domain adaptation [3]}, [4]}, [5]}, [6]}, [7]}, [8]} and few-shot/transfer learning [9]}, [10]}, [11]}, [12]}. Interestingly, we show that self-training can not only discover more objects, but also correct the initially noisy box labels. This result that neural networks can denoise noisy labels has been observed before [13]}, [14]}, [15]}, [16]}. Self-training also bears resemblance to other semi-supervised learning techniques[17]}, [18]}, [19]}, [20]} but is simpler and more broadly applicable.

[20]

https://openalex.org/W3001197829

02dc9dc1-6b79-4c4a-847c-2c5df7c8935b

Overview. We propose simple, high-level common-sense properties that can easily identify a few seed objects in the unlabeled data. These discovered objects then serve as labels to train an off-the-shelf object detector. Specifically, building upon the neural network's ability to learn consistent patterns from initial seed labels, we bootstrap the detector by self-training [1]}, [2]} using the same unlabeled data. The self-training process serves to correct and expand the initial pool of seed objects, gradually discovering more and more objects to further help train the detector. The whole process is summarized in alg:main. <FIGURE>

[2]

https://openalex.org/W3035160371

271449ee-de13-48fb-b9b5-b4e28235b0e9

What properties define mobile objects or traffic participants? Clearly, the most important characteristic is that they are mobile, i.e., they move around. If such an object is spotted at a particular location (e.g., a car at an intersection), it is unlikely that the object will still be there when one visits the intersection again a few days hence. In other words, mobile objects are ephemeral members of a scene [1]}. Of course, occasionally mobile objects like cars might be parked on the road for extended periods of time. However, for the most part ephemeral objects are likely to be mobile objects.

[1]

https://openalex.org/W2963170338

2788e5b9-e830-4e54-967a-21668742b9e4

We assume that our unlabeled data include a set of locations \(L\) which are traversed multiple times in separate driving sessions (or traversals). For every traversal \(t\) through location \(c \in L\) , we aggregate point clouds captured within a range of \([-H_s, H_e]\) of \(c\) to produce a dense 3D point cloud \(S_c^t\) for location \(c\) in traversal \(t\)We can easily transform captured point clouds to a shared coordinate frame via precise localization information through GPS/INS.. We then use these dense point clouds \(S_c^t\) to define ephemerality as described by Barnes et al. [1]}. Concretely, to check if a 3D point \({\mathbf {q}}\) in a scene is ephemeral, for each traversal \(t\) we can count the number \(N_t({\mathbf {q}})\) of LiDAR points that fall within a distance \(r\) to \({\mathbf {q}}\) , \(N_t \left(q\right) = \left|\left\lbrace p_i \mid \Vert p_i - q\Vert _{2} < r, p_i \in S_c^t \right\rbrace \right|.\)

[1]

https://openalex.org/W2963170338

4dd3e45d-4cb8-4c1e-9a3c-524275b03026

The graph structure together with the edge weights define a new metric that quantifies the similarity between two points. In this graph, two points that are connected by a path are considered to be close if the path has low total edge weight, namely, the points along the path share similar PP scores, indicating these points are likely from the same object. In contrast, a path in the graph that has high total edge weight likely goes across the boundary of two different objects (e.g., a mobile object and the background). Many graph-based clustering algorithms can fit the bill. We deploy the widely used DBSCAN [1]} algorithm for the clustering due to its simplicity and its ability to cluster without the need of setting the number of clusters beforehand. DBSCAN returns a list of clusters, from which we remove clusters of static (and thus persistent) background points by applying a threshold \(\gamma \) on the \(\alpha \) percentile of PP scores in a cluster (i.e., remove the cluster if the \(\alpha \) percentile of the PP scores in this cluster is larger than \(\gamma \) ). We then apply a straightforward bounding box fitting algorithm [2]} to fit an up-right bounding box to each cluster.

[1]

https://openalex.org/W1673310716

6201e7f7-3a1e-4090-8037-85dc3dc0b6ec

Concretely, we simply take off-the-shelf 3D object detectors [1]}, [2]}, [3]}, [4]} and directly train them from scratch on these initial seed labels via minimizing the corresponding detection loss from the detection algorithms.

[1]

https://openalex.org/W2949708697

8ae03edd-0665-494f-9ca8-ce09f486e586

Concretely, we simply take off-the-shelf 3D object detectors [1]}, [2]}, [3]}, [4]} and directly train them from scratch on these initial seed labels via minimizing the corresponding detection loss from the detection algorithms.

[3]

https://openalex.org/W2963727135

a0e21ac6-c0fe-49fe-a19c-15402908a7dd

Concretely, we simply take off-the-shelf 3D object detectors [1]}, [2]}, [3]}, [4]} and directly train them from scratch on these initial seed labels via minimizing the corresponding detection loss from the detection algorithms.

[4]

https://openalex.org/W2897529137

75f16918-8728-4067-aa02-bd9c5a5c93dc

Intriguingly, the object detector trained in this way outperforms the original seed bounding boxes themselves — the “detected” boxes have higher recall and are more accurate than the “discovered” boxes on the same training point clouds. See fig:teaser for an illustration. This phenomenon of a neural network improving on the provided noisy labels themselves is superficially surprising, but it has been observed before in other contexts [1]}. The key reason is that the noise is not consistent: the initial labels are generated scene-by-scene and object-by-object. In some scenes a car may be missed because it was parked throughout all traversals, while in many others it will be discovered as a mobile object. Even among discovered boxes of similar objects, some may miss a few foreground points but others wrongly include background points. The neural network, equipped with limited capacityWe note that for detection problems, a neural network can hardly over-fit the training data to achieve \(100\%\) accuracy [2]}, in contrast to classification problems [3]}., thus cannot reproduce this inconsistency and instead instead identifies the underlying consistent object patterns.

[1]

https://openalex.org/W2575671312

cb62e046-926a-4efe-9d5a-6a86e26b6b3c

Intriguingly, the object detector trained in this way outperforms the original seed bounding boxes themselves — the “detected” boxes have higher recall and are more accurate than the “discovered” boxes on the same training point clouds. See fig:teaser for an illustration. This phenomenon of a neural network improving on the provided noisy labels themselves is superficially surprising, but it has been observed before in other contexts [1]}. The key reason is that the noise is not consistent: the initial labels are generated scene-by-scene and object-by-object. In some scenes a car may be missed because it was parked throughout all traversals, while in many others it will be discovered as a mobile object. Even among discovered boxes of similar objects, some may miss a few foreground points but others wrongly include background points. The neural network, equipped with limited capacityWe note that for detection problems, a neural network can hardly over-fit the training data to achieve \(100\%\) accuracy [2]}, in contrast to classification problems [3]}., thus cannot reproduce this inconsistency and instead instead identifies the underlying consistent object patterns.

[2]

https://openalex.org/W3178826664

8a31145f-7c0d-4c50-8098-fac58f12af24

Intriguingly, the object detector trained in this way outperforms the original seed bounding boxes themselves — the “detected” boxes have higher recall and are more accurate than the “discovered” boxes on the same training point clouds. See fig:teaser for an illustration. This phenomenon of a neural network improving on the provided noisy labels themselves is superficially surprising, but it has been observed before in other contexts [1]}. The key reason is that the noise is not consistent: the initial labels are generated scene-by-scene and object-by-object. In some scenes a car may be missed because it was parked throughout all traversals, while in many others it will be discovered as a mobile object. Even among discovered boxes of similar objects, some may miss a few foreground points but others wrongly include background points. The neural network, equipped with limited capacityWe note that for detection problems, a neural network can hardly over-fit the training data to achieve \(100\%\) accuracy [2]}, in contrast to classification problems [3]}., thus cannot reproduce this inconsistency and instead instead identifies the underlying consistent object patterns.

[3]

https://openalex.org/W2566079294

d949639f-062f-40b8-8e4d-d6a178b412b8

Automatic improvement through self-training. Given that the trained detector has discovered many more objects, we can use the detector itself to produce an improved set of ground-truth labels, and re-train a new detector from scratch with these better ground truths. Furthermore, we can iterate this process: the new retrained detector has more positives and more accurate boxes for training, so it will likely have higher recall and better localization than the initial detector. As such we can use this second detector to produce a new set of pseudo-ground-truth boxes which can be used to train a third detector and so on. This iterative self-training [1]}, [2]} process will eventually converge when the pseudo-ground truth labeling is consistent with itself and the detector can no longer improve upon it.

[2]

https://openalex.org/W3035160371

2ec18b39-97e0-4dc6-9b28-4252a6728f02

Datasets. We validate our approach on two datasets: Lyft Level 5 Perception [1]} and nuScenes [2]}. To the best of our knowledge, these are the only two publicly available autonomous driving datasets that have both bounding box annotations and multiple traversals with accurate localization. To ensure fair assessment of generalizability, we re-split the dataset so that the training set and test set are geographically disjoint; we also discard locations with less than 2 examples in the training set. This results a train/test split of 11,873/4,901 point clouds for Lyft and 3,985/2,324 for nuScenes. To construct ground truth labels, we group all the traffic participants types in the original datasets into a single “mobile" object. Note that the ground-truth labels are only used for evaluation, not training.

[2]

https://openalex.org/W3035574168

d7e7d570-0746-4887-a62e-2309fe3d7b8e

In addition, we convert the raw Lyft and nuScenes data into the KITTI format to leverage off-the-shelf 3D object detectors that is predominantly built for KITTI  [1]}. We use the roof LiDAR (40 or 60 beam in Lyft; 32 beam in nuScenes), and the global 6-DoF localization along with the calibration matrices directly from the raw data.

[1]

https://openalex.org/W2115579991

d7b8a73b-26ca-4f3e-b681-ef94acdb7e7b

On localization. With current localization technology, we can reliably achieve accurate localization (e.g., 1-2 cm-level accuracy with RTKhttps://en.wikipedia.org/wiki/Real-time_kinematic_positioning, 10 cm-level with Monte Carlo Localization scheme [1]} as adopted in the nuScenes dataset [2]}). We assume good localization in the training set.

[2]

https://openalex.org/W3035574168

356886dd-fdee-4810-9ad2-25129961f7f1

Evaluation metric. We follow KITTI  [1]} to evaluate object detection in the bird's-eye view (BEV) and in 3D for the mobile objects. We report average precision (AP) with the intersection over union (IoU) thresholds at 0.5/0.25, which are used to evaluate cyclists and pedestrians objects in KITTI. We further follow [2]} to evaluate the AP at various depth ranges. Due to space constraints, we only present evaluation results with IoU=0.25 in tbl:main,tbl:ablation,tbl:nusc025,tbl:labelquality,tbl:kitti025,tbl:recallbyclassiou025summary. Please refer to the supplementary for results with IoU=0.5.

[1]

https://openalex.org/W2150066425

be7a2f02-59bb-4ed1-a2cd-567bade6990c

Evaluation metric. We follow KITTI  [1]} to evaluate object detection in the bird's-eye view (BEV) and in 3D for the mobile objects. We report average precision (AP) with the intersection over union (IoU) thresholds at 0.5/0.25, which are used to evaluate cyclists and pedestrians objects in KITTI. We further follow [2]} to evaluate the AP at various depth ranges. Due to space constraints, we only present evaluation results with IoU=0.25 in tbl:main,tbl:ablation,tbl:nusc025,tbl:labelquality,tbl:kitti025,tbl:recallbyclassiou025summary. Please refer to the supplementary for results with IoU=0.5.

[2]

https://openalex.org/W3034975685

cc48ad30-a6a7-46bf-b093-49186de7e8f9

Implementation. We present results on PointRCNN [1]} (the conclusions hold for other detectors such as PointPillars [2]}, and VoxelNet (SECOND) [3]}, [4]}. See more details in the supplementary materials). For reproducibility, we use the publicly available code from OpenPCDet [5]} for all models. We use the default hyperparameters tuned for KITTI except on the Lyft dataset in which we enlarge the perception range from 70m to 90m (since Lyft provides labels beyond 70m) and reduce the number of training epochs by 1/4 (since the training set is about three times of the size of KITTI). We default to 10 rounds of self-training (chosen arbitrarily due to compute constraints) and trained the model from scratch for each round of self-training. We also include results on PointRCNN trained up to 40 rounds, where we empirically observe that the performance converges (fig:ablationrounds). All models are trained with 4 NVIDIA 3090 GPUs. Please refer to the supplementary materials for full hyperparameters.

[1]

https://openalex.org/W2949708697

62d9cd11-7e91-4794-92de-4badebd15e3c

Implementation. We present results on PointRCNN [1]} (the conclusions hold for other detectors such as PointPillars [2]}, and VoxelNet (SECOND) [3]}, [4]}. See more details in the supplementary materials). For reproducibility, we use the publicly available code from OpenPCDet [5]} for all models. We use the default hyperparameters tuned for KITTI except on the Lyft dataset in which we enlarge the perception range from 70m to 90m (since Lyft provides labels beyond 70m) and reduce the number of training epochs by 1/4 (since the training set is about three times of the size of KITTI). We default to 10 rounds of self-training (chosen arbitrarily due to compute constraints) and trained the model from scratch for each round of self-training. We also include results on PointRCNN trained up to 40 rounds, where we empirically observe that the performance converges (fig:ablationrounds). All models are trained with 4 NVIDIA 3090 GPUs. Please refer to the supplementary materials for full hyperparameters.

[3]

https://openalex.org/W2963727135

9611e80e-ad1b-4737-895e-16a83762e127

Implementation. We present results on PointRCNN [1]} (the conclusions hold for other detectors such as PointPillars [2]}, and VoxelNet (SECOND) [3]}, [4]}. See more details in the supplementary materials). For reproducibility, we use the publicly available code from OpenPCDet [5]} for all models. We use the default hyperparameters tuned for KITTI except on the Lyft dataset in which we enlarge the perception range from 70m to 90m (since Lyft provides labels beyond 70m) and reduce the number of training epochs by 1/4 (since the training set is about three times of the size of KITTI). We default to 10 rounds of self-training (chosen arbitrarily due to compute constraints) and trained the model from scratch for each round of self-training. We also include results on PointRCNN trained up to 40 rounds, where we empirically observe that the performance converges (fig:ablationrounds). All models are trained with 4 NVIDIA 3090 GPUs. Please refer to the supplementary materials for full hyperparameters.

[4]

https://openalex.org/W2897529137

30512044-188f-464a-9a67-6499055ea4d1

Acknowledgements This research is supported by grants from the National Science Foundation NSF (III-1618134, III-1526012, IIS-1149882, IIS-1724282, TRIPODS-1740822, IIS-2107077, OAC-2118240, OAC-2112606 and IIS-2107161), the Office of Naval Research DOD (N00014-17-1-2175), the DARPA Learning with Less Labels program (HR001118S0044), the Bill and Melinda Gates Foundation, the Cornell Center for Materials Research with funding from the NSF MRSEC program (DMR-1719875), and SAP America. Supplementary Material Implementation details We set \([-H_s, H_e]\) to \([0, 70]\)  m since we experiment with frontal-view detection only. We combine only one scan into the dense point cloud \(S_c^t\) every 2 m within this range. In calculating PP score, we use as many traversals as possible (\(\ge 2\) ) and set \(r=0.3\) m. For clustering, we use \(K=70\) and \(r^{\prime }=2.0\) m in the graph, and \(\epsilon =0.1\) , min_samples \(=10\) for DBSCAN. For filtering, we use a loose threshold of \(\alpha =20\) percentile and \(\gamma =0.7\) . Other common sense properties are simply implemented as follows: [itemsep=1pt,topsep=3pt] \(\#\) points in the cluster \(>= 10\) ; Volume of fitted bounding boxes \(\in [0.5, 120]m^3\) ; The height (upright distance against the ground plane) of points \(\text{Height}_{\max } > 0.5m\) and \(\text{Height}_{\min } < 1.0m\) to ensure clusters not floating in the air or beneath the ground due to errors in LiDAR. We did not tune these parameters except qualitatively checked the fitted bounding boxes in few scenes in the Lyft “train” set. For detection models, we use the default hyper-parameters tuned on KITTIhttps://github.com/open-mmlab/OpenPCDet/tree/master/tools/cfgs/kitti_models with few exceptions listed in the paper. We will open-source the code upon acceptance. Experiments with other detectors Besides the PointRCNN detector [1]}, We experiment with two other detectors PointPillars [2]} and VoxelNet (SECOND) [3]}, [4]}, and show their results in tbl:second and tbl:pointpillars. We apply the default hyper-parameters of these two models tuned on KITTI, and apply the same procedure as that on PointRCNN models. Note that PointPillars and VoxelNet model need a pre-defined anchor size for different types of objects, which we picked (length, width, height) as (\(2.0, 1.0, 1.7\) ) m without tuning. We observe that generally the PointPillars and VoxelNet yield worse results than PointRCNN models (possibly due to the fixed anchor size for all mobile objects), but we still observe significant gains from self-training. <TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE> Detailed evaluation by object types In tbl:recallbyclass, we include detailed evaluations (BEV IoU\(=0.5\) , BEV IoU\(=0.25\) and by different depth ranges) of the recall of different object types in the Lyft test set. This corresponds to tbl:recallbyclassiou025summary in the main paper. Corresponding IoU=0.5 results We list the IoU=0.5 counterparts of tbl:main,tbl:ablation,tbl:nusc025,tbl:labelquality,tbl:kitti025,tbl:recallbyclassiou025summary in tbl:mainiou05,tbl:nusc05,tbl:kitti05,tbl:labelqualityiou05,tbl:ablation05. <FIGURE> Precision-recall evaluation In fig:ablationpr, we show how PR curve changes with different rounds of self-training: the max recall improves gradually while keeping high precision. This aligns with the expanded recall of the training set described above, and with what we observe qualitatively in fig:teaser. More qualitative results We show visualizations for additional qualitative results in fig:self-training for 5 additional LiDAR scenes. Visualizations show the progression of MODEST from seed generation, to detector trained on seed label set, to detector after 10 rounds of self training, and finally the ground truth bounding boxes. Observe that the detections obtain higher recall and learns a correct prior over object shapes as the method progresses. <FIGURE>

[1]

https://openalex.org/W2949708697

f98ceb15-6941-4259-a496-137e244149c4

Acknowledgements This research is supported by grants from the National Science Foundation NSF (III-1618134, III-1526012, IIS-1149882, IIS-1724282, TRIPODS-1740822, IIS-2107077, OAC-2118240, OAC-2112606 and IIS-2107161), the Office of Naval Research DOD (N00014-17-1-2175), the DARPA Learning with Less Labels program (HR001118S0044), the Bill and Melinda Gates Foundation, the Cornell Center for Materials Research with funding from the NSF MRSEC program (DMR-1719875), and SAP America. Supplementary Material Implementation details We set \([-H_s, H_e]\) to \([0, 70]\)  m since we experiment with frontal-view detection only. We combine only one scan into the dense point cloud \(S_c^t\) every 2 m within this range. In calculating PP score, we use as many traversals as possible (\(\ge 2\) ) and set \(r=0.3\) m. For clustering, we use \(K=70\) and \(r^{\prime }=2.0\) m in the graph, and \(\epsilon =0.1\) , min_samples \(=10\) for DBSCAN. For filtering, we use a loose threshold of \(\alpha =20\) percentile and \(\gamma =0.7\) . Other common sense properties are simply implemented as follows: [itemsep=1pt,topsep=3pt] \(\#\) points in the cluster \(>= 10\) ; Volume of fitted bounding boxes \(\in [0.5, 120]m^3\) ; The height (upright distance against the ground plane) of points \(\text{Height}_{\max } > 0.5m\) and \(\text{Height}_{\min } < 1.0m\) to ensure clusters not floating in the air or beneath the ground due to errors in LiDAR. We did not tune these parameters except qualitatively checked the fitted bounding boxes in few scenes in the Lyft “train” set. For detection models, we use the default hyper-parameters tuned on KITTIhttps://github.com/open-mmlab/OpenPCDet/tree/master/tools/cfgs/kitti_models with few exceptions listed in the paper. We will open-source the code upon acceptance. Experiments with other detectors Besides the PointRCNN detector [1]}, We experiment with two other detectors PointPillars [2]} and VoxelNet (SECOND) [3]}, [4]}, and show their results in tbl:second and tbl:pointpillars. We apply the default hyper-parameters of these two models tuned on KITTI, and apply the same procedure as that on PointRCNN models. Note that PointPillars and VoxelNet model need a pre-defined anchor size for different types of objects, which we picked (length, width, height) as (\(2.0, 1.0, 1.7\) ) m without tuning. We observe that generally the PointPillars and VoxelNet yield worse results than PointRCNN models (possibly due to the fixed anchor size for all mobile objects), but we still observe significant gains from self-training. <TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE> Detailed evaluation by object types In tbl:recallbyclass, we include detailed evaluations (BEV IoU\(=0.5\) , BEV IoU\(=0.25\) and by different depth ranges) of the recall of different object types in the Lyft test set. This corresponds to tbl:recallbyclassiou025summary in the main paper. Corresponding IoU=0.5 results We list the IoU=0.5 counterparts of tbl:main,tbl:ablation,tbl:nusc025,tbl:labelquality,tbl:kitti025,tbl:recallbyclassiou025summary in tbl:mainiou05,tbl:nusc05,tbl:kitti05,tbl:labelqualityiou05,tbl:ablation05. <FIGURE> Precision-recall evaluation In fig:ablationpr, we show how PR curve changes with different rounds of self-training: the max recall improves gradually while keeping high precision. This aligns with the expanded recall of the training set described above, and with what we observe qualitatively in fig:teaser. More qualitative results We show visualizations for additional qualitative results in fig:self-training for 5 additional LiDAR scenes. Visualizations show the progression of MODEST from seed generation, to detector trained on seed label set, to detector after 10 rounds of self training, and finally the ground truth bounding boxes. Observe that the detections obtain higher recall and learns a correct prior over object shapes as the method progresses. <FIGURE>

[3]

https://openalex.org/W2963727135

269e60a8-f66f-451c-8c4a-c206a34767bc

Acknowledgements This research is supported by grants from the National Science Foundation NSF (III-1618134, III-1526012, IIS-1149882, IIS-1724282, TRIPODS-1740822, IIS-2107077, OAC-2118240, OAC-2112606 and IIS-2107161), the Office of Naval Research DOD (N00014-17-1-2175), the DARPA Learning with Less Labels program (HR001118S0044), the Bill and Melinda Gates Foundation, the Cornell Center for Materials Research with funding from the NSF MRSEC program (DMR-1719875), and SAP America. Supplementary Material Implementation details We set \([-H_s, H_e]\) to \([0, 70]\)  m since we experiment with frontal-view detection only. We combine only one scan into the dense point cloud \(S_c^t\) every 2 m within this range. In calculating PP score, we use as many traversals as possible (\(\ge 2\) ) and set \(r=0.3\) m. For clustering, we use \(K=70\) and \(r^{\prime }=2.0\) m in the graph, and \(\epsilon =0.1\) , min_samples \(=10\) for DBSCAN. For filtering, we use a loose threshold of \(\alpha =20\) percentile and \(\gamma =0.7\) . Other common sense properties are simply implemented as follows: [itemsep=1pt,topsep=3pt] \(\#\) points in the cluster \(>= 10\) ; Volume of fitted bounding boxes \(\in [0.5, 120]m^3\) ; The height (upright distance against the ground plane) of points \(\text{Height}_{\max } > 0.5m\) and \(\text{Height}_{\min } < 1.0m\) to ensure clusters not floating in the air or beneath the ground due to errors in LiDAR. We did not tune these parameters except qualitatively checked the fitted bounding boxes in few scenes in the Lyft “train” set. For detection models, we use the default hyper-parameters tuned on KITTIhttps://github.com/open-mmlab/OpenPCDet/tree/master/tools/cfgs/kitti_models with few exceptions listed in the paper. We will open-source the code upon acceptance. Experiments with other detectors Besides the PointRCNN detector [1]}, We experiment with two other detectors PointPillars [2]} and VoxelNet (SECOND) [3]}, [4]}, and show their results in tbl:second and tbl:pointpillars. We apply the default hyper-parameters of these two models tuned on KITTI, and apply the same procedure as that on PointRCNN models. Note that PointPillars and VoxelNet model need a pre-defined anchor size for different types of objects, which we picked (length, width, height) as (\(2.0, 1.0, 1.7\) ) m without tuning. We observe that generally the PointPillars and VoxelNet yield worse results than PointRCNN models (possibly due to the fixed anchor size for all mobile objects), but we still observe significant gains from self-training. <TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE> Detailed evaluation by object types In tbl:recallbyclass, we include detailed evaluations (BEV IoU\(=0.5\) , BEV IoU\(=0.25\) and by different depth ranges) of the recall of different object types in the Lyft test set. This corresponds to tbl:recallbyclassiou025summary in the main paper. Corresponding IoU=0.5 results We list the IoU=0.5 counterparts of tbl:main,tbl:ablation,tbl:nusc025,tbl:labelquality,tbl:kitti025,tbl:recallbyclassiou025summary in tbl:mainiou05,tbl:nusc05,tbl:kitti05,tbl:labelqualityiou05,tbl:ablation05. <FIGURE> Precision-recall evaluation In fig:ablationpr, we show how PR curve changes with different rounds of self-training: the max recall improves gradually while keeping high precision. This aligns with the expanded recall of the training set described above, and with what we observe qualitatively in fig:teaser. More qualitative results We show visualizations for additional qualitative results in fig:self-training for 5 additional LiDAR scenes. Visualizations show the progression of MODEST from seed generation, to detector trained on seed label set, to detector after 10 rounds of self training, and finally the ground truth bounding boxes. Observe that the detections obtain higher recall and learns a correct prior over object shapes as the method progresses. <FIGURE>

[4]

https://openalex.org/W2897529137

91a91ad7-c57f-46bf-851d-513f26cb4e83

Besides the PointRCNN detector [1]}, We experiment with two other detectors PointPillars [2]} and VoxelNet (SECOND) [3]}, [4]}, and show their results in tbl:second and tbl:pointpillars. We apply the default hyper-parameters of these two models tuned on KITTI, and apply the same procedure as that on PointRCNN models. Note that PointPillars and VoxelNet model need a pre-defined anchor size for different types of objects, which we picked (length, width, height) as (\(2.0, 1.0, 1.7\) ) m without tuning. We observe that generally the PointPillars and VoxelNet yield worse results than PointRCNN models (possibly due to the fixed anchor size for all mobile objects), but we still observe significant gains from self-training. <TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE>

[1]

https://openalex.org/W2949708697

2e44d642-1e6c-465c-99c3-b4c2e600c56f

Besides the PointRCNN detector [1]}, We experiment with two other detectors PointPillars [2]} and VoxelNet (SECOND) [3]}, [4]}, and show their results in tbl:second and tbl:pointpillars. We apply the default hyper-parameters of these two models tuned on KITTI, and apply the same procedure as that on PointRCNN models. Note that PointPillars and VoxelNet model need a pre-defined anchor size for different types of objects, which we picked (length, width, height) as (\(2.0, 1.0, 1.7\) ) m without tuning. We observe that generally the PointPillars and VoxelNet yield worse results than PointRCNN models (possibly due to the fixed anchor size for all mobile objects), but we still observe significant gains from self-training. <TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE>

[3]

https://openalex.org/W2963727135

4e47c945-7464-492c-af5f-fbf0e8bbccb8

Besides the PointRCNN detector [1]}, We experiment with two other detectors PointPillars [2]} and VoxelNet (SECOND) [3]}, [4]}, and show their results in tbl:second and tbl:pointpillars. We apply the default hyper-parameters of these two models tuned on KITTI, and apply the same procedure as that on PointRCNN models. Note that PointPillars and VoxelNet model need a pre-defined anchor size for different types of objects, which we picked (length, width, height) as (\(2.0, 1.0, 1.7\) ) m without tuning. We observe that generally the PointPillars and VoxelNet yield worse results than PointRCNN models (possibly due to the fixed anchor size for all mobile objects), but we still observe significant gains from self-training. <TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE><TABLE>

[4]

https://openalex.org/W2897529137

a30e2945-44da-4aa7-b69c-9008f8ae8ab3

Image inpainting, or image completion, is a task about image synthesis technique aims to filling occluded regions or missing pixels with appropriate semantic contents. The main objective of image inpainting is producing visually authentic images with less semantic inconsistency using computer vision-based approaches. Traditional methods relied on a patch-based matching approach using the measurement of cosine similarity [1]}. Recently, the remarkable capability of generative adversarial networks (GAN) [2]} has boosted image inpainting performance based on convolutional neural networks (CNN). Because of its hierarchical design, GAN with encoder-decoder structure has exceptional reconstruction ability compared to previous approaches. The decoder synthesizes visual images from the feature level as the encoder learns how to extract feature representations from images. Currently, GAN-based approaches constitute a dominant stream in image inpainting [3]}, [4]}, [5]}, [6]}, [7]}, [8]}.

[3]

https://openalex.org/W2963420272

d08be01e-5527-45bb-8962-682e875615bd

Image inpainting, or image completion, is a task about image synthesis technique aims to filling occluded regions or missing pixels with appropriate semantic contents. The main objective of image inpainting is producing visually authentic images with less semantic inconsistency using computer vision-based approaches. Traditional methods relied on a patch-based matching approach using the measurement of cosine similarity [1]}. Recently, the remarkable capability of generative adversarial networks (GAN) [2]} has boosted image inpainting performance based on convolutional neural networks (CNN). Because of its hierarchical design, GAN with encoder-decoder structure has exceptional reconstruction ability compared to previous approaches. The decoder synthesizes visual images from the feature level as the encoder learns how to extract feature representations from images. Currently, GAN-based approaches constitute a dominant stream in image inpainting [3]}, [4]}, [5]}, [6]}, [7]}, [8]}.

[4]

https://openalex.org/W2738588019

f725bb2c-14c4-40de-97cc-7e7a1a7eefeb

Image inpainting, or image completion, is a task about image synthesis technique aims to filling occluded regions or missing pixels with appropriate semantic contents. The main objective of image inpainting is producing visually authentic images with less semantic inconsistency using computer vision-based approaches. Traditional methods relied on a patch-based matching approach using the measurement of cosine similarity [1]}. Recently, the remarkable capability of generative adversarial networks (GAN) [2]} has boosted image inpainting performance based on convolutional neural networks (CNN). Because of its hierarchical design, GAN with encoder-decoder structure has exceptional reconstruction ability compared to previous approaches. The decoder synthesizes visual images from the feature level as the encoder learns how to extract feature representations from images. Currently, GAN-based approaches constitute a dominant stream in image inpainting [3]}, [4]}, [5]}, [6]}, [7]}, [8]}.

[5]

https://openalex.org/W2796286534

4aac10d4-3934-4eeb-9c20-8ed4d5ef875f

Image inpainting, or image completion, is a task about image synthesis technique aims to filling occluded regions or missing pixels with appropriate semantic contents. The main objective of image inpainting is producing visually authentic images with less semantic inconsistency using computer vision-based approaches. Traditional methods relied on a patch-based matching approach using the measurement of cosine similarity [1]}. Recently, the remarkable capability of generative adversarial networks (GAN) [2]} has boosted image inpainting performance based on convolutional neural networks (CNN). Because of its hierarchical design, GAN with encoder-decoder structure has exceptional reconstruction ability compared to previous approaches. The decoder synthesizes visual images from the feature level as the encoder learns how to extract feature representations from images. Currently, GAN-based approaches constitute a dominant stream in image inpainting [3]}, [4]}, [5]}, [6]}, [7]}, [8]}.

[6]

https://openalex.org/W3043547428

833eac53-d21d-48d1-8556-63ab46e85b9e

Image inpainting, or image completion, is a task about image synthesis technique aims to filling occluded regions or missing pixels with appropriate semantic contents. The main objective of image inpainting is producing visually authentic images with less semantic inconsistency using computer vision-based approaches. Traditional methods relied on a patch-based matching approach using the measurement of cosine similarity [1]}. Recently, the remarkable capability of generative adversarial networks (GAN) [2]} has boosted image inpainting performance based on convolutional neural networks (CNN). Because of its hierarchical design, GAN with encoder-decoder structure has exceptional reconstruction ability compared to previous approaches. The decoder synthesizes visual images from the feature level as the encoder learns how to extract feature representations from images. Currently, GAN-based approaches constitute a dominant stream in image inpainting [3]}, [4]}, [5]}, [6]}, [7]}, [8]}.

[7]

https://openalex.org/W2982763192

f0d84e4a-e011-43aa-86b7-3861411d01da

Image inpainting, or image completion, is a task about image synthesis technique aims to filling occluded regions or missing pixels with appropriate semantic contents. The main objective of image inpainting is producing visually authentic images with less semantic inconsistency using computer vision-based approaches. Traditional methods relied on a patch-based matching approach using the measurement of cosine similarity [1]}. Recently, the remarkable capability of generative adversarial networks (GAN) [2]} has boosted image inpainting performance based on convolutional neural networks (CNN). Because of its hierarchical design, GAN with encoder-decoder structure has exceptional reconstruction ability compared to previous approaches. The decoder synthesizes visual images from the feature level as the encoder learns how to extract feature representations from images. Currently, GAN-based approaches constitute a dominant stream in image inpainting [3]}, [4]}, [5]}, [6]}, [7]}, [8]}.

[8]

https://openalex.org/W3026446890

64304234-ed2b-40fd-9f92-ba48976da5e8

However, despite GAN's high image restoration performance, some pixel artifacts or color inconsistency called 'fake texture' inevitably occur in the process of decoding [1]}, [2]}. Fake pixels cause degradation of image restoration performance by dropping the appearance consistency in the synthesized image. To tackle this issue, we introduce dynamic attention map (DAM) that detects fake textures in feature map and highlights them by generating an attention mask (or attention map) [3]} for image inpainting. Unlike existing GAN-based inpainting methods requiring high computational cost for generating attention map [4]}, [5]}, our proposed DAM blocks exploit learnable convolutional layers for detecting fake texture and converting it into an attention map for each different scale of each decoding layer. We reported the comparisons on CelebA-HQ and Places2 datasets and showed that outcome of our DAM-GAN demonstrating higher quality than other existing inpainting methods including GAN-based approaches. <FIGURE>

[2]

https://openalex.org/W3012472557

91f8c837-28f1-4f41-b55c-057a1c9dd28b

However, despite GAN's high image restoration performance, some pixel artifacts or color inconsistency called 'fake texture' inevitably occur in the process of decoding [1]}, [2]}. Fake pixels cause degradation of image restoration performance by dropping the appearance consistency in the synthesized image. To tackle this issue, we introduce dynamic attention map (DAM) that detects fake textures in feature map and highlights them by generating an attention mask (or attention map) [3]} for image inpainting. Unlike existing GAN-based inpainting methods requiring high computational cost for generating attention map [4]}, [5]}, our proposed DAM blocks exploit learnable convolutional layers for detecting fake texture and converting it into an attention map for each different scale of each decoding layer. We reported the comparisons on CelebA-HQ and Places2 datasets and showed that outcome of our DAM-GAN demonstrating higher quality than other existing inpainting methods including GAN-based approaches. <FIGURE>

[3]

https://openalex.org/W2804078698

5f76847c-cf08-4fb3-ab5a-895ab8592911

However, despite GAN's high image restoration performance, some pixel artifacts or color inconsistency called 'fake texture' inevitably occur in the process of decoding [1]}, [2]}. Fake pixels cause degradation of image restoration performance by dropping the appearance consistency in the synthesized image. To tackle this issue, we introduce dynamic attention map (DAM) that detects fake textures in feature map and highlights them by generating an attention mask (or attention map) [3]} for image inpainting. Unlike existing GAN-based inpainting methods requiring high computational cost for generating attention map [4]}, [5]}, our proposed DAM blocks exploit learnable convolutional layers for detecting fake texture and converting it into an attention map for each different scale of each decoding layer. We reported the comparisons on CelebA-HQ and Places2 datasets and showed that outcome of our DAM-GAN demonstrating higher quality than other existing inpainting methods including GAN-based approaches. <FIGURE>

[4]

https://openalex.org/W2985764327

206db2dd-0956-4b1b-be7f-ff78ef22c107

However, despite GAN's high image restoration performance, some pixel artifacts or color inconsistency called 'fake texture' inevitably occur in the process of decoding [1]}, [2]}. Fake pixels cause degradation of image restoration performance by dropping the appearance consistency in the synthesized image. To tackle this issue, we introduce dynamic attention map (DAM) that detects fake textures in feature map and highlights them by generating an attention mask (or attention map) [3]} for image inpainting. Unlike existing GAN-based inpainting methods requiring high computational cost for generating attention map [4]}, [5]}, our proposed DAM blocks exploit learnable convolutional layers for detecting fake texture and converting it into an attention map for each different scale of each decoding layer. We reported the comparisons on CelebA-HQ and Places2 datasets and showed that outcome of our DAM-GAN demonstrating higher quality than other existing inpainting methods including GAN-based approaches. <FIGURE>

[5]

https://openalex.org/W3026446890

64d69187-7a1a-4281-8336-e682ab994d65

Traditional image inpainting methods were based on the exemplar-search approach, which divides image into patches to refill missing areas with other patches according to similarity computations such as PatchMatch [1]}. Recently, progressive improvement of deep learning based generative models have demonstrated high feasibility for image synthesis. Especially GAN [2]} demonstrates brilliant performance in image inpainting tasks. Context Encoders (CE) [3]} adopted encoder-decoder based GAN for image inpainting and Globally and Locally (GL) [4]} incorporates global and local generators to maintain pixel consistency of output images. Contextual Attention (CA) [5]} imitated the traditional patch-based method using GAN to take advantage of the basic concept of conventional exemplar-based methods. However, CE [3]}, GL [4]} and CA [5]} have limitations on refilling irregular regions because of their local region based discriminators. Since they are usually specialized in reconstructing rectangular masks, images with free-shaped masks will decrease the quality of outputs. To tackle this limitations, recent inpainting approaches tend to remove local discriminator on architecture [9]}.

[3]

https://openalex.org/W2963420272

7b18da8d-5f3e-4c13-af18-0a7143549d35

Traditional image inpainting methods were based on the exemplar-search approach, which divides image into patches to refill missing areas with other patches according to similarity computations such as PatchMatch [1]}. Recently, progressive improvement of deep learning based generative models have demonstrated high feasibility for image synthesis. Especially GAN [2]} demonstrates brilliant performance in image inpainting tasks. Context Encoders (CE) [3]} adopted encoder-decoder based GAN for image inpainting and Globally and Locally (GL) [4]} incorporates global and local generators to maintain pixel consistency of output images. Contextual Attention (CA) [5]} imitated the traditional patch-based method using GAN to take advantage of the basic concept of conventional exemplar-based methods. However, CE [3]}, GL [4]} and CA [5]} have limitations on refilling irregular regions because of their local region based discriminators. Since they are usually specialized in reconstructing rectangular masks, images with free-shaped masks will decrease the quality of outputs. To tackle this limitations, recent inpainting approaches tend to remove local discriminator on architecture [9]}.

[4]

https://openalex.org/W2738588019

f63a1da9-3bee-4629-be9a-b3f0969fcbc9

Traditional image inpainting methods were based on the exemplar-search approach, which divides image into patches to refill missing areas with other patches according to similarity computations such as PatchMatch [1]}. Recently, progressive improvement of deep learning based generative models have demonstrated high feasibility for image synthesis. Especially GAN [2]} demonstrates brilliant performance in image inpainting tasks. Context Encoders (CE) [3]} adopted encoder-decoder based GAN for image inpainting and Globally and Locally (GL) [4]} incorporates global and local generators to maintain pixel consistency of output images. Contextual Attention (CA) [5]} imitated the traditional patch-based method using GAN to take advantage of the basic concept of conventional exemplar-based methods. However, CE [3]}, GL [4]} and CA [5]} have limitations on refilling irregular regions because of their local region based discriminators. Since they are usually specialized in reconstructing rectangular masks, images with free-shaped masks will decrease the quality of outputs. To tackle this limitations, recent inpainting approaches tend to remove local discriminator on architecture [9]}.

[5]

https://openalex.org/W3043547428

c8b04b9f-4157-4698-83dc-124e9b9502bc

Partial conv [1]} did not employ GAN for inpainting, but solved the problem of generalization on irregular masks. It propose rule-based binary mask which is updated layer by layer in encoder-decoder network and showed high feasibility of refilling irregular masks. This mask-based inpainting approach is advanced in Gated conv [2]} by adopting GAN and replacing rule-based mask with learnable mask. Both Partial conv [1]} and Gated conv [2]} put forward a mask-based weights map for feature maps in the decoding process, similar to attention map [5]} based method.

[1]

https://openalex.org/W2798365772

61a1fb55-80c9-4bbf-adf8-1209a3e80e9c

Partial conv [1]} did not employ GAN for inpainting, but solved the problem of generalization on irregular masks. It propose rule-based binary mask which is updated layer by layer in encoder-decoder network and showed high feasibility of refilling irregular masks. This mask-based inpainting approach is advanced in Gated conv [2]} by adopting GAN and replacing rule-based mask with learnable mask. Both Partial conv [1]} and Gated conv [2]} put forward a mask-based weights map for feature maps in the decoding process, similar to attention map [5]} based method.

[2]

https://openalex.org/W2982763192

311a7cb3-6c20-4889-8f40-e6f5a14039b6

Partial conv [1]} did not employ GAN for inpainting, but solved the problem of generalization on irregular masks. It propose rule-based binary mask which is updated layer by layer in encoder-decoder network and showed high feasibility of refilling irregular masks. This mask-based inpainting approach is advanced in Gated conv [2]} by adopting GAN and replacing rule-based mask with learnable mask. Both Partial conv [1]} and Gated conv [2]} put forward a mask-based weights map for feature maps in the decoding process, similar to attention map [5]} based method.

[5]

https://openalex.org/W2804078698

2487d61a-b68b-42ed-8e2a-a145fcf52944

The goal of generator \(G\) is to fill missing parts with appropriate contents by understanding the input image \(x\) (encoding) and synthesizing the output image \(G(x)\) (decoding). Fig. REF describes the overall architecture of generator \(G\) . The coarse reconstruction stage begins by filling pixels with a rough texture. The DAM reconstruction then uses DAM blocks to restore the coarse output \({G}_{C}(x)\) with detailed contents. We defined the residual convolution layer by combining residual block [1]} and convolution layer, and we adopted concatenation-based skip-connection [2]} and dilated convolution [3]} in the middle of the generator. Skip-connections have a notable effect on reducing vanishing gradient problems and maintaining spatial information of reconstructed images, and dilated convolution increases the receptive field to enhance the efficiency of the computations.

[1]

https://openalex.org/W2194775991

589e2785-8407-4215-96bd-9d561d799c2b

The goal of generator \(G\) is to fill missing parts with appropriate contents by understanding the input image \(x\) (encoding) and synthesizing the output image \(G(x)\) (decoding). Fig. REF describes the overall architecture of generator \(G\) . The coarse reconstruction stage begins by filling pixels with a rough texture. The DAM reconstruction then uses DAM blocks to restore the coarse output \({G}_{C}(x)\) with detailed contents. We defined the residual convolution layer by combining residual block [1]} and convolution layer, and we adopted concatenation-based skip-connection [2]} and dilated convolution [3]} in the middle of the generator. Skip-connections have a notable effect on reducing vanishing gradient problems and maintaining spatial information of reconstructed images, and dilated convolution increases the receptive field to enhance the efficiency of the computations.

[2]

https://openalex.org/W1901129140

635a17e8-f207-4696-8943-e86350a5d9ab

The goal of generator \(G\) is to fill missing parts with appropriate contents by understanding the input image \(x\) (encoding) and synthesizing the output image \(G(x)\) (decoding). Fig. REF describes the overall architecture of generator \(G\) . The coarse reconstruction stage begins by filling pixels with a rough texture. The DAM reconstruction then uses DAM blocks to restore the coarse output \({G}_{C}(x)\) with detailed contents. We defined the residual convolution layer by combining residual block [1]} and convolution layer, and we adopted concatenation-based skip-connection [2]} and dilated convolution [3]} in the middle of the generator. Skip-connections have a notable effect on reducing vanishing gradient problems and maintaining spatial information of reconstructed images, and dilated convolution increases the receptive field to enhance the efficiency of the computations.

[3]

https://openalex.org/W2412782625

7106f6f6-e877-49cd-a447-5a11cdbcfcca

Discriminator \(D\) serves as a criticizer that distinguishes between real and synthesized images. Adversarial training between \(G\) and \(D\) can further improve the quality of synthesized image. Because local discriminator has critical limitations on handling irregular mask as mentioned in section 2., we use one global discriminator for adversarial training our model. We employed the global discriminator from CA [1]}.

[1]

https://openalex.org/W3043547428

2d314833-f297-4066-9ce0-7d5924329d78

Similar to fakeness prediction in [1]}, fakeness map \({M}_{i}\) is produced through 1x1 convolutional filters and sigmoid function from feature \({F}_{i}\) . Then, we can use \({M}_{i}\) as an attention map like [2]}. After element-wise multiplication of \({M}_{i} \otimes {F}_{i}\) , the output feature \({F^{\prime }}_{i}\) is obtained. Then element-wise sum \({F}_{i} \oplus {F^{\prime }}_{i}\) becomes the final output \({T}_{i-1}\) , which is upsampled and passed to the upper layer in the decoder. Fakeness map \({M}_{i}\) is trainable dynamically in each layer from decoder using DAM loss \(\mathcal {L}_{DAM}\) , which is expressed in section 3.

[2]

https://openalex.org/W2804078698

89439ebe-58df-4219-9ad6-ff5934c03876

Our model was trained on two datasets: CelebA-HQ and [1]} Places2 [2]}. We randomly divided the 30,000 images in CelebA-HQ dataset into a training set of 27,000 images and a validation set of 3,000 images. In Places2 dataset, we select same categories as [3]} in training set and tested our model on validation set. All images are resized to 128 \(\times \) 128.

[1]

https://openalex.org/W2962760235

12a97c61-29f3-46b3-af9c-4f5747670de4

Our model was trained on two datasets: CelebA-HQ and [1]} Places2 [2]}. We randomly divided the 30,000 images in CelebA-HQ dataset into a training set of 27,000 images and a validation set of 3,000 images. In Places2 dataset, we select same categories as [3]} in training set and tested our model on validation set. All images are resized to 128 \(\times \) 128.

[2]

https://openalex.org/W2732026016

694fed95-3aaa-45ae-9911-6fb09a6cebc8

Our model was trained on two datasets: CelebA-HQ and [1]} Places2 [2]}. We randomly divided the 30,000 images in CelebA-HQ dataset into a training set of 27,000 images and a validation set of 3,000 images. In Places2 dataset, we select same categories as [3]} in training set and tested our model on validation set. All images are resized to 128 \(\times \) 128.

[3]

https://openalex.org/W3175375202

cd10e4d4-0796-4156-b537-a717d2b1c4f2

To prepare input images for our model, we defined the centered mask and random mask. The centered mask has 64 \(\times \) 64 size fixed in the center of the image, and the random mask has an irregular shape following the mask generation approach in [1]}. We used an ADAM optimizer [2]} in this experiment, and hyper-parameters are set to \({\lambda }_{re}=1, {\lambda }_{adv}=0.001\) and \({\lambda }_{DAM}=0.005\) .

[2]

https://openalex.org/W2964121744