As Hulk only used open-source datasets, we provide the links to the datasets used in the project. For some datasets that need to be preprocessed, we also provide the code to preprocess them or upload the proper preprocessed annotation files.
We sort the datasets by their tasks as follows:
- Human parsing
- 2D pose estimation
- Attribute recognition
- Pedestrian detection
- Skeleton-based action recognition
- Pedestrian image caption
- 3D pose estimation & Mesh recovery
After downloading and preprocessing the datasets, do NOT forget to put them to the relative place referring the training yaml.