How to improve the PoseTrack results

[ericwang0701]

Thank you for sharing such amazing work especially integrating different Pose Tracking Modules into AlphaPose.

I have tested the existing 3 Pose Track modules (i.e. Human-ReID based, Detector based, and PoseFlow ) on custom videos. But I noticed that human tracking will become unreliable when the people are occluded from each other or when their feet are out of camera view. Here is one example, you can see that the Pose Track keep detecting a "new person" (i.e. the skeleton color changes)

However, just like the demo videos in this Github project. The Poses Track modules seem should work significantly better even in the more complicated multi-person videos. Any suggestion on how to improve the PoseTrack results would be very helpful. Thanks

[Huhaowen0130]

I have met similar problems...Have you solved it already? I guess it may be refined if trained with own dataset.

[sjwang222]

Hi Eric,

I met the similar issue. Since in your case humans to be detected do not change their dress, maybe you can try to extract the colors (RGB values) around the shoulders or hips to re-identify humans in the video. Note that results.json stores the absolute coordinates for 26 joints and human index for each frame. It is viable to extract RGB values around joints and adjust the human index correspondingly (and finally re-use their rendering modular to obtain newly detected videos).