Add Guided Anchoring.

[learncrazy]

🚀 Feature

Recently I saw a blog and wrote that there is a way to predict the shape and position of Anchor directly through Guided Anchoring without clustering Anchor based on data, and it can already be applied to yolov3 to improve mAP. So I think yolov5 can try to add this.

https://arxiv.org/pdf/1901.03278.pdf

Motivation

Pitch

Alternatives

Additional context

[glenn-jocher]

@learncrazy I don't have time myself to experiment with this, but if you can obtain improved results with this method we can help you work on a PR.

I skimmed the paper very quickly, and the premise seems to be to exploit a priori information about correlations in the 4d (xywh) box space, though they seem to mainly discuss applicability to 2-stage detectors.

We already harvest and illustrate some of this information to present to the user in labels.png. For example the VOC dataset:

[glenn-jocher]

I created a correlogram of the 4 box dimensions using seaborn. The existing plots in labels.png are circled below. There are definitely constraints that exist in the 4d space. I believe these are already being enforced however, by clipping out of bounds boxes in test.py and detect.py.

A more sophisticated solution may employ a 4d density estimate to apply a bayesian prior onto the network output. This might reduce confidences in low density areas for example. This could even be done by class for high statistics datasets like COCO.

[glenn-jocher]

Added automatic correlogram plotting if user has seaborn and pandas installed 987c2268490a6220ac14a1325e76774c60ea4f47

[learncrazy]

Of course, I think your work is excellent. If I can, I will try it and see if mAP can be improved. Thanks for your answer.

[buimanhlinh96]

@learncrazy Did you have any experiment with this idea and how about the result?

[github-actions[bot]]

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.