Hugging Face's logo

Datasets:
byeonghwikim
/
abp_dataset

License:
Dataset card Files Community
Dataset Viewer
Full Screen
The dataset could not be loaded because the splits use different data file formats, which is not supported. Read more about the splits configuration. Click for more details.
Couldn't infer the same data file format for all splits. Got {NamedSplit('train'): ('json', {}), NamedSplit('validation'): (None, {})}
Error code:   FileFormatMismatchBetweenSplitsError

Need help to make the dataset viewer work? Make sure to review how to configure the dataset viewer, and open a discussion for direct support.

ALFRED Dataset for ABP

We provide the ALFRED dataset used for ABP including ResNet-18 features of egocentric and surrounding views, annotations, etc. The surrdounding views are from four navigable actions defined in ALFRED: RotateLeft (90°), LookUp(15°), LookDown(15°), and RotateRight(90°). The file structure is almost identical to the ALFRED dataset, so refer to ALFRED for more details.

Download the dataset

Raw RGB images with depth and object masks

Move to the root (denoted by ALFRED_ROOT below) of the ABP (or related work) repo and clone this repository by following the commands below.

git clone https://huggingface.co/datasets/byeonghwikim/abp_images json_2.1.0

Pre-extracted features

To skip feature extraction from RGB images, you can use our pre-extracted features. Move to the root (denoted by ALFRED_ROOT below) of the ABP (or related work) repo and clone this repository by following the commands below.
Note: This dataset is quite large (~1.6T).

cd $ALFRED_ROOT/data
git clone https://huggingface.co/datasets/byeonghwikim/abp_dataset json_feat_2.1.0

After downloading the dataset, you may directly load a surrounding feature and the expected outcome is as below.

>> import torch
>> filename = 'train/look_at_obj_in_light-AlarmClock-None-DeskLamp-301/trial_T20190907_174127_043461/feat_conv_panoramic.pt'
>> im = torch.load(filename) # [5, T, 512, 7, 7], T the length of a trajectory
>> im.shape
torch.Size([5, T, 512, 7, 7])

The 0-dimension of the feature corresponds to the respective view directions as below.

  • 0: left view (RotateLeft)
  • 1: up view (LookUp)
  • 2: front (egocentric) view (no action)
  • 3: down view (LookDown)
  • 4: right view (RotateRight)

Inspired by MOCA, we apply image augmentation to the agent's visual observation. We apply two types of image augmentation: 1) swapping color channels of images and 2) AutoAugment.

  • No augmentation: (feat_conv_panoramic.pt)
  • Swapping color channels: (feat_conv_colorSwap1_panoramic.pt, feat_conv_colorSwap2_panoramic.pt)
  • AutoAugment: (feat_conv_onlyAutoAug1_panoramic.pt ~ feat_conv_onlyAutoAug4_panoramic.pt)

Related work that uses this dataset

Citation

If you find this repository useful, please cite this repository.

@inproceedings{kim2021agent,
  author    = {Kim, Byeonghwi and Bhambri, Suvaansh and Singh, Kunal Pratap and Mottaghi, Roozbeh and Choi, Jonghyun},
  title     = {Agent with the Big Picture: Perceiving Surroundings for Interactive Instruction Following},
  booktitle = {Embodied AI Workshop @ CVPR 2021},
  year      = {2021},
}
Downloads last month
62
Edit dataset card