Will the RFLA degrade the performance on large objects ?

[Icecream-blue-sky]

I find that only the performance on small objects is showed in Table 7. What about the performance on medium and large objects (i.e., AP_m, AP_l) ?

[Icecream-blue-sky]

Since the RFLA is proposed to deal with the scene where the anchor and gt do not overlap, which can also be solved by DIoU and CIoU, how does RFLA perform compared to DIoU and CIoU ? Thanks!

[Chasel-Tsui]

Hi, there will be a slight degradation on large objects since the k in the assigner is too small to detect large objects. As for the DIoU and CIoU based assigning, I am not sure of their performance but the Gaussian modelling can pay more attention to the center region compared to them. But if it is combined with a ranking-based assigner, intuitively, it may yield a similar result as the gaussian modelling since they can also measure the similarity of outliers.

[Icecream-blue-sky]

Have you tried setting different k for different sizes of gt? I don’t think it is necessary to use the same k for all objects, for example YOLOX sets the k value according to IoU.

[Chasel-Tsui]

Yes, you are right, setting different k for different sized gt will be much better for generic object detection. For multi-scale object detection, k should be adjusted according to its scale. But if we focus on tiny object detection, where there are many tiny objects, it is might be better to use the same k to compensate some positive samples for tiny objects. Tiny objects usually yield a low IoU with anchor or get a low IoU prediction when estimating the dynamic k, resulting in unbalanced training and lack of samples for them. I agree that the fixed k is at the cost of the degradation of large objects. I also think it is a problem worth studying to find the trade-off between sample quality and quantity/to obtain balanced training without the loss of samples for large objects.

[Icecream-blue-sky]

Thank you for your kind reply！Nice work!