What is the purpose of anchor_t?

[Uoops]

❔Question

Thank you for sharing your work! It is the best YOLO related repo in python I have ever seen!

Can you explain what anchor_t means and what it does in the training process?

The reason is that I notice you use anchor_t instead of iou_t to generate the targets compared to your YOLOv3 work. I think iou_t is for the classification, can anchor_t does the same thing?
# Matches
r = t[:, :, 4:6] / anchors[:, None] # wh ratio
j = torch.max(r, 1. / r).max(2)[0] < model.hyp['anchor_t'] # compare
# j = wh_iou(anchors, t[:, 4:6]) > model.hyp['iou_t'] # iou(3,n)=wh_iou(anchors(3,2), gwh(n,2))
t = t[j] # filter

Additional context

The code above is copied from yolov5/utils/general.py line 565

[github-actions]

Hello @Uoops, thank you for your interest in our work! Please visit our Custom Training Tutorial to get started, and see our Jupyter Notebook , Docker Image, and Google Cloud Quickstart Guide for example environments.

If this is a bug report, please provide screenshots and minimum viable code to reproduce your issue, otherwise we can not help you.

If this is a custom model or data training question, please note Ultralytics does not provide free personal support. As a leader in vision ML and AI, we do offer professional consulting, from simple expert advice up to delivery of fully customized, end-to-end production solutions for our clients, such as:

• Cloud-based AI systems operating on hundreds of HD video streams in realtime.
• Edge AI integrated into custom iOS and Android apps for realtime 30 FPS video inference.
• Custom data training, hyperparameter evolution, and model exportation to any destination.

For more information please visit https://www.ultralytics.com.

[glenn-jocher]

@Uoops yes, we use anchor_t instead of iou_t in YOLOv5. These are both thresholds for deciding which labels to match to which anchors.

Using anchor_t now provides certainty that anchors are always capable of fitting a given label, and also allows you to program the range of the width-height outputs to most efficiently occupy the width-height space.

[GorillaSX]

hey @glenn-jocher , thanks for your answer for the purpose of anchor_t. But I kind of don't understand the logic and algorithm behind anchor_t, so would you mind give me a hint or share the algorithm name you're using here? I believe a good understanding of anchor_t, could help us to set right threshold here. I appreciate it.

[GorillaSX]

hey @glenn-jocher , I think I figured it out.

[glenn-jocher]

@GorillaSX anchor_t is the anchor width and height multiple threshold used to select label-anchor matches when computing loss.

[GorillaSX]

hey @glenn-jocher thanks for your answer. I understood that. What confused me is those two lines code,

 r = t[:, :, 4:6] / anchors[:, None]  # wh ratio
 j = torch.max(r, 1. / r).max(2)[0] < model.hyp['anchor_t']  # compar

But I think I figured it out at this time, still thanks.

[glenn-jocher]

j are target-anchor match candidates.

[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.

[lfydegithub]

@GorillaSX anchor_t is the anchor width and height multiple threshold used to select label-anchor matches when computing loss.

thank you for your reply, so how to get my dataset anchor_t threshold ? appreciate it