thanks Nik!

Two main points in this PR:

• The first one is that I think with the code as is, the ground truth path is in practice a required argument. But we want to be able to run the detector+tracker even if no ground truth is provided (i.e., running inference). If ground truth is provided, we run inference and evaluation, and optionally (if save_video_output is True) we can save a video with detections and ground truth.

• The second one is about the frame_number vs frame_idx. Apologies cause I think I may have not been very clear before. The issue is that we shouldn't be using frame numbers as indices to a list, because the frame numbers are extracted from the png filenames, and these may not necessarily start with frame 1 or 0. The safest, most general way imo is to use dictionaries whose keys are frame numbers (not indices), for both predictions and ground truth.

  At the moment, we use dictionaries with keys = frame numbers for ground truth only. For predictions, the keys are frame indices.

  I believe this would cause issues if the ground truth frames start counting from a number different from 1 - I include a specific example in the comments.

I think also some bits from previous PRs got lost? I found some old versions of lines (e.g. line 224 of crabs/tracker/utils/tracking.py ).

Let me know if this makes sense, happy to discuss.

Thanks Sofia. Sorry for the messy code. I have now add code to deal when we have no gt_path. Sorry about the frame_idx vs frame_number. It was my fault for not checking that, just the bad way of merging happened.
I wasn't try to be a rebellious! 😂