Firstly, thank you for your excellent work. When using these released video data and labels, I have met two key troubles. Hope you can give me some help and suggestions.
How to get the accurate bounding boxes of persons automatically form labels? I'm now generating the approximate bounding boxes from their reprojected 2D keypoints. However, this way is not good enough for supervised learning.
Can we get the visibility of 2D keypoints about persons in still frames? Pose estimaion benckmarks including COCO, MPII and CrowdPose all provide the visibility of joints (v=0 when it is missed or not annotated, v=1 when it is occluded, v=2 when it is visiable). After processing file hdPose3d_stage1_coco19.tar, I can only get two types (v=0 and v=2). So, how to judge one joint whether it is occluded or not from current labels?
cucdengjunli
commented
1 year ago
Firstly, thank you for your excellent work. When using these released video data and labels, I have met two key troubles. Hope you can give me some help and suggestions.
hdPose3d_stage1_coco19.tar, I can only get two types (v=0 and v=2). So, how to judge one joint whether it is occluded or not from current labels?