Skip to content

Latest commit

 

History

History
29 lines (21 loc) · 2.34 KB

DATASET.md

File metadata and controls

29 lines (21 loc) · 2.34 KB

Overall pipeline

The dataset files should be located at data/. The paths of the video/images(frames) and train/val splits should be defined in dataet_configs/[dataset_name].py. The dataloader is defined in exp_configs/[dataset_name]/[model_name]-[experiment_name].py.

HMDB-51

  1. Download HMDB51 and the official three splits here.
  2. Use submodules/video_datasets_api/hmdb_tools/hmdb_extract_rar.sh to extract the rar file.
  3. Use submodules/video_datasets_api/hmdb_tools/hmdb_extract_frames.sh to extract videos into frames. (optional, it depends on the dataloader you use)
    a. Because of what is defined in dataset_configs/hmdb.py, the frames have to be located at `data/hmdb51/frames_q5'
  4. Use tools/datasets/hmdb_splits_to_csv_frame_extracted.sh to convert the official three splits to the format required for this package.
    a. Because of what is defined in dataset_configs/hmdb.py, the splits have to be located at `data/hmdb51/splits_frames'

Refer to examples/01-Prepare_HMDB51.ipynb for details.

Something-Something-V1

  1. Go to the TwentyBN's download page. Copy the HTML source code and make save it as TwentyBN.html
  2. Use submodules/video_datasets_api/something-something-v1_tools/download_sthsth.py --version 1 TwentyBN.html [download_dir] to download the dataset.
  3. Extract using cat [download_dir]/20bn-something-something-v1-?? | tar zx -C data/something-something-v1/frames.
  4. Download annotations and save into data/something-something-v1/annotations
  5. Use python tools/datasets/generate_somethingv1_splits.py --root data/something-something-v1/frames data/something-something-v1/splits_frames data/something-something-v1/annotations to generate splits required to use the PyVideoAI dataloaders.

CATER (Task 1/2, static/cameramotion)

  1. Use submodules/video_datasets_api/cater_tools/download_cater_max2action.sh [download_dir] to download and extract both max2action and max2action_cameramotion.
  2. Put the dataset in data/cater (make symbolic link)
  3. Extract into frames using submodules/video_datasets_api/cater_tools/cater_extract_frames.sh.
  4. Use python tools/datasets/generate_cater_task1_2_splits.py to generate training splits.