COCO-WholeBody — The First Dataset to Evaluate The Whole Body Pose

Source: Arxiv

COCO-WholeBody is the first dataset for evaluating whole body posture. COCO-WholeBody is an extension of the COCO 2017 dataset with the same training and validation breakdowns as COCO. There are 4 types of object boundaries available for each person: person box, face box, left-hand box, and right-hand box. In addition, 133 key points: 17 for the body, 6 for the legs, 68 for the face, and 42 for the arms. The dataset is available exclusively for research purposes. Commercial use is prohibited.

An example of marking points of a body. Source: Arxiv

How the dataset was collected

The marking process consisted of the following steps:

1. For each person, the boundaries of the face, left hand and right hand were manually marked;
2. Quality assessment: a separate group of markers checked the quality of the lines;
3. For each marked face and hand, a pre-trained model was used to mark key points. HRNetV2 was used as the detector architecture;
4. The predicted markup was then manually checked. After that, the quality of manual marking was checked separately.