README.md

metadata

dataset_info:
  features:
    - name: image
      dtype: image
    - name: file_name
      dtype: string
    - name: keypoints
      list:
        list: float32
        length: 2
      length: 18
  splits:
    - name: train
      num_bytes: 443109457
      num_examples: 5000
    - name: test
      num_bytes: 26950250
      num_examples: 300
  download_size: 470171720
  dataset_size: 470059707
configs:
  - config_name: default
    data_files:
      - split: train
        path: data/train-*
      - split: test
        path: data/test-*

A subset of Taichi-HD dataset with estimated body landmarks, ported from https://github.com/AliaksandrSiarohin/motion-cosegmentation.

1. The protocol of dataset creation can be found at the original paper: https://arxiv.org/abs/2004.03234
2. We backup the raw images (./raw_assets) and landmarks (./landmark) in this repo.