Detectron2 Model Zoo and Baselines

Introduction

This file documents a large collection of baselines trained with detectron2 in Sep-Oct, 2019. All numbers were obtained on Big Basin servers with 8 NVIDIA V100 GPUs & NVLink. The speed numbers are periodically updated with latest PyTorch/CUDA/cuDNN versions. You can access these models from code using detectron2.model_zoo APIs.

In addition to these official baseline models, you can find more models in projects/.

How to Read the Tables

The "Name" column contains a link to the config file. Models can be reproduced using tools/train_net.py with the corresponding yaml config file, or tools/lazyconfig_train_net.py for python config files.
Training speed is averaged across the entire training. We keep updating the speed with latest version of detectron2/pytorch/etc., so they might be different from the metrics file. Training speed for multi-machine jobs is not provided.
Inference speed is measured by tools/train_net.py --eval-only, or inference_on_dataset(), with batch size 1 in detectron2 directly. Measuring it with custom code may introduce other overhead. Actual deployment in production should in general be faster than the given inference speed due to more optimizations.
The model id column is provided for ease of reference. To check downloaded file integrity, any model on this page contains its md5 prefix in its file name.
Training curves and other statistics can be found in metrics for each model.

Common Settings for COCO Models

All COCO models were trained on train2017 and evaluated on val2017.
The default settings are not directly comparable with Detectron's standard settings. For example, our default training data augmentation uses scale jittering in addition to horizontal flipping.

To make fair comparisons with Detectron's settings, see Detectron1-Comparisons for accuracy comparison, and benchmarks for speed comparison.
For Faster/Mask R-CNN, we provide baselines based on 3 different backbone combinations:
- FPN: Use a ResNet+FPN backbone with standard conv and FC heads for mask and box prediction, respectively. It obtains the best speed/accuracy tradeoff, but the other two are still useful for research.
- C4: Use a ResNet conv4 backbone with conv5 head. The original baseline in the Faster R-CNN paper.
- DC5 (Dilated-C5): Use a ResNet conv5 backbone with dilations in conv5, and standard conv and FC heads for mask and box prediction, respectively. This is used by the Deformable ConvNet paper.
Most models are trained with the 3x schedule (~~37 COCO epochs). Although 1x models are heavily under-trained, we provide some ResNet-50 models with the 1x (~~12 COCO epochs) training schedule for comparison when doing quick research iteration.

ImageNet Pretrained Models

It's common to initialize from backbone models pre-trained on ImageNet classification tasks. The following backbone models are available:

R-50.pkl: converted copy of MSRA's original ResNet-50 model.
R-101.pkl: converted copy of MSRA's original ResNet-101 model.
X-101-32x8d.pkl: ResNeXt-101-32x8d model trained with Caffe2 at FB.
R-50.pkl (torchvision): converted copy of torchvision's ResNet-50 model. More details can be found in the conversion script.

Note that the above models have different format from those provided in Detectron: we do not fuse BatchNorm into an affine layer. Pretrained models in Detectron's format can still be used. For example:

X-152-32x8d-IN5k.pkl: ResNeXt-152-32x8d model trained on ImageNet-5k with Caffe2 at FB (see ResNeXt paper for details on ImageNet-5k).
R-50-GN.pkl: ResNet-50 with Group Normalization.
R-101-GN.pkl: ResNet-101 with Group Normalization.

These models require slightly different settings regarding normalization and architecture. See the model zoo configs for reference.

License

All models available for download through this document are licensed under the Creative Commons Attribution-ShareAlike 3.0 license.

COCO Object Detection Baselines

Faster R-CNN:

Name	lr sched	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	model id	download
R50-C4	1x	0.551	0.102	4.8	35.7	137257644	model \| metrics
R50-DC5	1x	0.380	0.068	5.0	37.3	137847829	model \| metrics
R50-FPN	1x	0.210	0.038	3.0	37.9	137257794	model \| metrics
R50-C4	3x	0.543	0.104	4.8	38.4	137849393	model \| metrics
R50-DC5	3x	0.378	0.070	5.0	39.0	137849425	model \| metrics
R50-FPN	3x	0.209	0.038	3.0	40.2	137849458	model \| metrics
R101-C4	3x	0.619	0.139	5.9	41.1	138204752	model \| metrics
R101-DC5	3x	0.452	0.086	6.1	40.6	138204841	model \| metrics
R101-FPN	3x	0.286	0.051	4.1	42.0	137851257	model \| metrics
X101-FPN	3x	0.638	0.098	6.7	43.0	139173657	model \| metrics

RetinaNet:

Name	lr sched	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	model id	download
R50	1x	0.205	0.041	4.1	37.4	190397773	model \| metrics
R50	3x	0.205	0.041	4.1	38.7	190397829	model \| metrics
R101	3x	0.291	0.054	5.2	40.4	190397697	model \| metrics

RPN & Fast R-CNN:

Name	lr sched	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	prop. AR	model id	download
RPN R50-C4	1x	0.130	0.034	1.5		51.6	137258005	model \| metrics
RPN R50-FPN	1x	0.186	0.032	2.7		58.0	137258492	model \| metrics
Fast R-CNN R50-FPN	1x	0.140	0.029	2.6	37.8		137635226	model \| metrics

COCO Instance Segmentation Baselines with Mask R-CNN

Name	lr sched	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	mask AP	model id	download
R50-C4	1x	0.584	0.110	5.2	36.8	32.2	137259246	model \| metrics
R50-DC5	1x	0.471	0.076	6.5	38.3	34.2	137260150	model \| metrics
R50-FPN	1x	0.261	0.043	3.4	38.6	35.2	137260431	model \| metrics
R50-C4	3x	0.575	0.111	5.2	39.8	34.4	137849525	model \| metrics
R50-DC5	3x	0.470	0.076	6.5	40.0	35.9	137849551	model \| metrics
R50-FPN	3x	0.261	0.043	3.4	41.0	37.2	137849600	model \| metrics
R101-C4	3x	0.652	0.145	6.3	42.6	36.7	138363239	model \| metrics
R101-DC5	3x	0.545	0.092	7.6	41.9	37.3	138363294	model \| metrics
R101-FPN	3x	0.340	0.056	4.6	42.9	38.6	138205316	model \| metrics
X101-FPN	3x	0.690	0.103	7.2	44.3	39.5	139653917	model \| metrics

New baselines using Large-Scale Jitter and Longer Training Schedule

The following baselines of COCO Instance Segmentation with Mask R-CNN are generated using a longer training schedule and large-scale jitter as described in Google's Simple Copy-Paste Data Augmentation paper. These models are trained from scratch using random initialization. These baselines exceed the previous Mask R-CNN baselines.

In the following table, one epoch consists of training on 118000 COCO images.

Name	epochs	train time (s/im)	inference time (s/im)	box AP	mask AP	model id	download
R50-FPN	100	0.376	0.069	44.6	40.3	42047764	model \| metrics
R50-FPN	200	0.376	0.069	46.3	41.7	42047638	model \| metrics
R50-FPN	400	0.376	0.069	47.4	42.5	42019571	model \| metrics
R101-FPN	100	0.518	0.073	46.4	41.6	42025812	model \| metrics
R101-FPN	200	0.518	0.073	48.0	43.1	42131867	model \| metrics
R101-FPN	400	0.518	0.073	48.9	43.7	42073830	model \| metrics
regnetx_4gf_dds_FPN	100	0.474	0.071	46.0	41.3	42047771	model \| metrics
regnetx_4gf_dds_FPN	200	0.474	0.071	48.1	43.1	42132721	model \| metrics
regnetx_4gf_dds_FPN	400	0.474	0.071	48.6	43.5	42025447	model \| metrics
regnety_4gf_dds_FPN	100	0.487	0.073	46.1	41.6	42047784	model \| metrics
regnety_4gf_dds_FPN	200	0.487	0.072	47.8	43.0	42047642	model \| metrics
regnety_4gf_dds_FPN	400	0.487	0.072	48.2	43.3	42045954	model \| metrics

COCO Person Keypoint Detection Baselines with Keypoint R-CNN

Name	lr sched	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	kp. AP	model id	download
R50-FPN	1x	0.315	0.072	5.0	53.6	64.0	137261548	model \| metrics
R50-FPN	3x	0.316	0.066	5.0	55.4	65.5	137849621	model \| metrics
R101-FPN	3x	0.390	0.076	6.1	56.4	66.1	138363331	model \| metrics
X101-FPN	3x	0.738	0.121	8.7	57.3	66.0	139686956	model \| metrics

COCO Panoptic Segmentation Baselines with Panoptic FPN

Name	lr sched	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	mask AP	PQ	model id	download
R50-FPN	1x	0.304	0.053	4.8	37.6	34.7	39.4	139514544	model \| metrics
R50-FPN	3x	0.302	0.053	4.8	40.0	36.5	41.5	139514569	model \| metrics
R101-FPN	3x	0.392	0.066	6.0	42.4	38.5	43.0	139514519	model \| metrics

LVIS Instance Segmentation Baselines with Mask R-CNN

Mask R-CNN baselines on the LVIS dataset, v0.5. These baselines are described in Table 3(c) of the LVIS paper.

NOTE: the 1x schedule here has the same amount of iterations as the COCO 1x baselines. They are roughly 24 epochs of LVISv0.5 data. The final results of these configs have large variance across different runs.

Name	lr sched	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	mask AP	model id	download
R50-FPN	1x	0.292	0.107	7.1	23.6	24.4	144219072	model \| metrics
R101-FPN	1x	0.371	0.114	7.8	25.6	25.9	144219035	model \| metrics
X101-FPN	1x	0.712	0.151	10.2	26.7	27.1	144219108	model \| metrics

Cityscapes & Pascal VOC Baselines

Simple baselines for

Mask R-CNN on Cityscapes instance segmentation (initialized from COCO pre-training, then trained on Cityscapes fine annotations only)
Faster R-CNN on PASCAL VOC object detection (trained on VOC 2007 train+val + VOC 2012 train+val, tested on VOC 2007 using 11-point interpolated AP)

Name	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	box AP50	mask AP	model id	download
R50-FPN, Cityscapes	0.240	0.078	4.4			36.5	142423278	model \| metrics
R50-C4, VOC	0.537	0.081	4.8	51.9	80.3		142202221	model \| metrics

Other Settings

Ablations for Deformable Conv and Cascade R-CNN:

Name	lr sched	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	mask AP	model id	download
Baseline R50-FPN	1x	0.261	0.043	3.4	38.6	35.2	137260431	model \| metrics
Deformable Conv	1x	0.342	0.048	3.5	41.5	37.5	138602867	model \| metrics
Cascade R-CNN	1x	0.317	0.052	4.0	42.1	36.4	138602847	model \| metrics
Baseline R50-FPN	3x	0.261	0.043	3.4	41.0	37.2	137849600	model \| metrics
Deformable Conv	3x	0.349	0.047	3.5	42.7	38.5	144998336	model \| metrics
Cascade R-CNN	3x	0.328	0.053	4.0	44.3	38.5	144998488	model \| metrics

Ablations for normalization methods, and a few models trained from scratch following Rethinking ImageNet Pre-training. (Note: The baseline uses 2fc head while the others use 4conv1fc head)

Name	lr sched	train time (s/iter)	inference time (s/im)	train mem (GB)	box AP	mask AP	model id	download
Baseline R50-FPN	3x	0.261	0.043	3.4	41.0	37.2	137849600	model \| metrics
GN	3x	0.309	0.060	5.6	42.6	38.6	138602888	model \| metrics
SyncBN	3x	0.345	0.053	5.5	41.9	37.8	169527823	model \| metrics
GN (from scratch)	3x	0.338	0.061	7.2	39.9	36.6	138602908	model \| metrics
GN (from scratch)	9x	N/A	0.061	7.2	43.7	39.6	183808979	model \| metrics
SyncBN (from scratch)	9x	N/A	0.055	7.2	43.6	39.3	184226666	model \| metrics

A few very large models trained for a long time, for demo purposes. They are trained using multiple machines:

Name	inference time (s/im)	train mem (GB)	box AP	mask AP	PQ	model id	download
Panoptic FPN R101	0.098	11.4	47.4	41.3	46.1	139797668	model \| metrics
Mask R-CNN X152	0.234	15.1	50.2	44.0		18131413	model \| metrics
above + test-time aug.			51.9	45.9