for video inference I am missing most inference on most frames even though viz is good

[vv111y]

I am modifying the demo code for top down heatmap inference demo/top_down_pose_tracking_demo_with_mmdet.py in order to extract the keypoints for each frame for further processing. However, most predictions coming out of either the model or tracking call are empty even though the output video has correct keypoints for each video frame. I am at a loss what is happening - how do I get the keypoint predictions per frame?

I have looked through the code and am at model.show_result.
It doesn't make sense that predictions are reused, i have advanced through the video frame by frame and there are clear changes for every frame.

number of predictions: 512
number of frames (and model calls): 5272

mmpose v0.24.0

thank you,