DATASET.md

DATASET

I. Our custom dataset format

All data should be converted to our custom dataset format before being used for training. Our format has this folder structure:

dataset_name/
    images/
    train.json
    val.json
    test.json

• images is a folder containing image files.
• train.json, val.json, test.json are annotation files. Here are an example of labels in these files:

[
    {
        "image": "001.png",
        "points": [[280, 540], [315, 468], [356, 354], [354, 243], [471, 331], [514, 440], [546, 540]],
        "visibility": [1, 1, 1, 1, 0, 0, 1]
    }
    {
        "image": "002.png",
        "points": [[269, 529], [289, 465], [305, 410], [310, 309], [455, 358], [542, 429], [560, 542]],
        "visibility": [1, 0, 0, 1, 1, 1, 1]
    },
    ...
]

II. LSP and LSPET

• Link to LSP dataset: https://sam.johnson.io/research/lsp.html.
• Link to LSPET dataset: https://sam.johnson.io/research/lspet.html.

1. Convert annotation to JSON format

• The annotation contains x and y locations and a binary value indicating the visbility of joints.
• Use tools/lsp_data_to_json.py to convert LSP and LSPET annotation files to json format:
• NOTE: We removed 6061 images from LSPET dataset due to missing points.

python tools/lsp_data_to_json.py --image_folder=data/lsp_dataset/images --input_file data/lsp_dataset/joints.mat --output_file data/lsp_dataset/labels.json
python tools/lsp_data_to_json.py --image_folder=data/lspet_dataset/images --input_file data/lspet_dataset/joints.mat --output_file data/lspet_dataset/labels.json

2. Merge 2 dataset and divide into subsets

• Training: 3739 from LSPET and 1800 from LSP.
• Validation: 100 from LSPET and 100 from LSP.
• Test: 100 from LSPET and 100 from LSP.

Please update paths to LSP and LSPET in tools/split_lsp_lspet.py and run:

python tools/split_lsp_lspet.py

III. MPII Humanpose

• We only use images with numOtherPeople = 0. The original dataset are divided into 3 subsets:

• Training: 9503 images.
• Validation: 1000 images.
• Test: 1000 images.

IV. PushUp dataset

We have push-up 420 videos, divided in 3 sets:

• Training: 8837 images from 317 videos.
• Validation: 1189 images from 41 videos.
• Test: 1013 images from 62 videos.