Not a real pose classification

[syan-dev]

Hi, thank you very much for publishing the code and model weight. I am trying to understand your code in inference_and_visualization.py and am having some difficulties.

mae = abs(gt_count - pose_count) / (gt_count + 1e-9)
if mae < best_mae:
          best_mae = mae
          real_action = action_type
          real_index = index

So gt_count is a predefined variable from the dataset representing the ground truth count. In your code, you are trying to count repetitions for different classes such as push-ups, pull-ups, bench presses, etc. You calculate the difference between your count and the ground truth count and select the count that is closest to the ground truth.

If I understand correctly, you are counting on the ground truth known in advance. I tried to delete gt_count and your code no longer works.

[hitcbw]

Same question. The method is class-agnostic and the best result will be preserved as the final result.

[syan-dev]

@hitcbw To be honest, I tried so many repositories and still found that RepNet provides a stable and light solution. Would you happen to have any other better approach?