The Uniform Matching part of the paper and the code are inconsistent

[hhaAndroid]

Very good job, but I don't know why The Uniform Matching part of the paper and the code are inconsistent?

paper:

1. adopting the k nearest anchor as positive anchors for each ground-truth box
2. set IoU thresholds in Uniform Matching to ignore large IoU (>0:7) negative anchors and small IoU (<0.15) positive anchors

code:

# positive indices when matching predict boxes and gt boxes
        indices = [
            tuple(
                torch.topk(
                    c[i],
                    k=self.match_times,
                    dim=0,
                    largest=False)[1].numpy().tolist()
            )
            for i, c in enumerate(C.split(sizes, -1))
        ]
 # positive indices when matching anchor boxes and gt boxes
        indices1 = [
            tuple(
                torch.topk(
                    c[i],
                    k=self.match_times,
                    dim=0,
                    largest=False)[1].numpy().tolist())
            for i, c in enumerate(C1.split(sizes, -1))]

Looking forward to your answer！

[chensnathan]

Hi, thanks for your interest.

In the uniform_matcher, we mark the indexes of both topk anchors and topk predict boxes as positives, which gives stable and slightly higher performance than only use the topk anchors(37.7 vs. 37.1).

We implement the code for thresholds in the loss function. You can find it here(L344-L365).

[hhaAndroid]

Hi. thank you for your reply. Thanks, I get it. I think this step is more critical. Have you considered updating the paper about uniform_matcher?

[chensnathan]

Thanks for your advice. Although I don't agree with the idea that the step is more critical, I will update the paper with this part.

[hhaAndroid]

Thank you for your reply. Looking forward to your future work！