Open tucan9389 opened 5 years ago
It's need gaussian filter after output and then average the coordinates between the n-frames for reducing shake in real-time inference.
Indeed, it resulted low precision in keypoints regression.
I still working on it...
Hey, hi again! I'm feedbacked about needs to post process on Android project. In my case, PoseEstimation-CoreML, I'm not any post process after inference. Just pick the x, y indices of maximum value from the heatmaps result.
1. Is there any need to post process after inference?
2. Besides, is it related to #61 issue(output resolution is only 24x24)?
Thank you.