NVIDIA-AI-IOT
ParseObjects for openpose original output
Hi! Thank you for your incredible lib!
I've been trying use your ParseObjects for original openpose output, and for paf it doesn't work, while for cmap peaks are found correctly.
For topology I used POSE_BODY_PART_PAIRS from /openpose/src/openpose/pose/poseParameters.cpp.
It seems like I have to do some postprocessing for paf, but it's not obvious.
Please can you help me with this?
I have faced a similar problem. Tried to use the lib on heatmaps and pafs, produced by openpose_pytorch model. Used the topology for COCO in the same file(witch matches the topology in openpose_pytorch sources), but got poor body parts connections(though good peaks were found). Is there any additional constraint on paf predictions in your model comparing to the original openpose? How exactly the skeleton vector in human_pose.json should be formed?
