This PR fixes a bug that introduces a mismatch between label-coordinates and video, when extracting cropped outlier frames.

bug description
When using cropping, the predicted keypoints produced with analyze_videos are stored in crop-window coordinates. This is also reflected in the metadata.pickle. In the current implementation of extract_outlier_frames, key-point coordinates are ALWAYS converted back to full-frame coordinates by adding the crop margin read from the metadata pickle file. (introduced in PR #2538)

This should not always the intended behavior: e.g. when the user tries to extract cropped output frames for a next refinement iteration, the keypoints should also reflect cropped coordinates.

See this imageforum issue and potentially related github issues: #3060, #3130, #1979

Steps to reproduce:

1. Create a new project with labeled data and train a network.
2. Run analyze_video with cropping enabled, e.g. :

deeplabcut.analyze_videos(
    config_path,
    [video_path],
    shuffle=shuffle,
    trainingsetindex=trainingsetindex,
    cropping=[908, 1920, 0, 1080],
)

3. Add the analyzed video to the video sets in config.yaml and specify cropping coordinates there.

video_sets: 
   path/to/video.mp4:
      crop: 908, 1920, 0, 1080

4. Run extract_outlier_frames:

deeplabcut.extract_outlier_frames(
    config_path,
    [video_path],
    shuffle=shuffle,
    trainingsetindex=trainingsetindex,
    automatic=True,
)

5. In Napari, open the created image folder inside labeled-frames to check the labels.

Omitting step 3:

Including step3:

After accepting proposed code change in this PR:

changed behavior:
extract_outlier_frames reads the fields video_sets in config and checks if cropping is configured. The keypoints are converted by the difference between margins in metadata.pickle and the output crop margin specified in video_sets, instead of always adding the margins specified in the metadata.pickle.