DATASET.md

Dataset Preparation

MPII-Cooking-2 Dataset

The MPII-Cooking-2 dataset and can be downloaded from here. After downloading the dataset, the directory structure should look like this:

.
├── experimentalSetup
├── videos
│   ├── s07-d72-cam-002.avi
│   └── ...
└── attributesAnnotations_MPII-Cooking-2.mat

We have provided the mpii_split.py script to split the dataset into training, validation, and testing set. The script will generate the following directory structure:

MPII
├── train.csv
├── test.csv
├── train
│   ├── addV
│   │   ├── s22-d26-cam-002_8666.mp4
│   │   ├── s22-d26-cam-002_8711.mp4
│   │   └── ...
|   ├── arrangeV
│   │   ├── s28-d51-cam-002_2860.mp4
│   │   ├── s28-d51-cam-002_3142.mp4
│   │   └── ...
│   └── ...
└── test
    ├── addV
    │   ├── s08-d02-cam-002_5200.mp4
    │   ├── s08-d02-cam-002_5260.mp4
    │   └── ...
    └── ...

The original dataloader is modified to support the hdf5 format (in this demo we skip the validation set for simplicity). We have prepared the code in the convert_h5.py script to convert the dataset to hdf5 format, where the final directory structure should look like the same as the original dataset structure but in one hdf5 file.

Other Datasets

To construct the dataset in hdf5 format

1. After donwloading the dataset, prepare the csv files for training, validation, and testing set as train.csv, val.csv, test.csv. The format of the csv file is:

   path_to_video_1 label_1
   path_to_video_2 label_2
   path_to_video_3 label_3
   ...
   path_to_video_N label_N
   

2. Please put all annotation json files and the frame lists or videos under the same directory, then refer to the convert_h5.py script to convert the dataset to hdf5 format as the MPII-Cooking-2 dataset preparation above.

3. Set DATA.PATH_TO_DATA_DIR to the appropriate path.

4. Set DATA.PATH_PREFIX to be the path to the folder containing extracted frames accordingly.