Documentation on aspect ratio threshold

[ml5ah]

🚀 Feature

Hi @glenn-jocher
Thanks for the amazing work with yolov5
This feature is related to documentation on the aspect ratio threshold in utils/datasets.py that is used for custom training.

Motivation

I have been working on object detection projects and using YOLOv5 frequently for some of the training tasks at hand. However, most of the time, mAP performance did not match my expectations set seeing the mAP scores on the COCO dataset. My dataset has mostly rectangular objects, and about 10-15% of them being either very wide or very tall. Going through one of the issues on GitHub, I realized that there is an aspect ratio threshold that discards all objects having an aspect ratio > 20 by default.

Pitch

It would be great to make this clear in the train custom dataset (https://github.com/ultralytics/yolov5/wiki/Train-Custom-Data)
so that we can adjust this parameter from the get go during experimentation. Thanks for the help!

[glenn-jocher]

@ml5ah thanks for the feedback! Yes perhaps box candidate thresholds should be included in the hyperparameters used for training to make them easier to modify and easier to spot. Can you submit a PR for this to help others? I will add a TODO here to update the Custom Training Tutorial with this information.

In general you should also be aware that the default architecture is not optimized for high aspect ratio objects, and the models would benefit from changes targeted specifically to this domain, which may include elongation of the convolution kernels in the C3 layers, i.e. perhaps alternate 5x1 and 1x5 convolutions in place of, or in addition to, the current 3x3 and 1x1 sequences.

TODO: Update Train Custom Data Tutorial with label candidate criteria section, as well as NMS settings section.

[cszer]

Very good note 👍👍

[github-actions]

This issue has been automatically marked as stale because it has not had recent activity. It will be closed if no further activity occurs. Thank you for your contributions.

[MrinalJain17]

@glenn-jocher Can we revive this issue? I'd be happy to draft a PR.

[ml5ah]

@MrinalJain17 could yo please review this PR? #2869
@glenn-jocher thank you!

[MrinalJain17]

@ml5ah Oh, I didn't realize there was already a PR on this. Thanks of the info. Hope this feature gets merged soon.

[glenn-jocher]

TODO removed, default AR threshold increased to 100 in #5556