Hand keypoints are wrong in SDK 3.7.7 body tracking sample code for BODY_FORMAT::POSE_34


I installed the SDK (v. 3.7.7 with CUDA 11.7) on Ubuntu 20.04 and ran the sample code for body tracking, where only the following lines were adjusted (tested with ZED 2 and Nvidia RTX 3090):

obj_param.detection_model = sl.DETECTION_MODEL.HUMAN_BODY_ACCURATE
obj_param.body_format = sl.BODY_FORMAT.POSE_34 

Unfortunately, the keypoints of the face and hands are not detected correctly at all. Especially the hand/thumb keypoints are very important for our use cases.

In the following, an example is provided.

We plan to use multiple ZED cameras and are disappointed that this promoted feature does not work out of the box. We hope that there is a fix to this issue.

Thank you!

PS.: As a new user, I am not allowed to upload more than one media item.


In the current version, we lack precision on the hands and feets. These are interpolated from the ankle and the wrist, and no further rotation is apply. In next versions, we’ll add AI that also detects hands and feets, much more accurately.

About the face, you can use the person head detection, it will be better than the human body detection.