Binhui Xie (谢斌辉)

thanks!

Sorry, another question. https://github.com/IDEACVR/DINO/blob/67bbcd97ef30a48cf343b7b0f3ad9ea0795b6fcd/models/dino/dino.py#L384 Why is `sigmoid_focal_loss` (binary cross entropy with logits) usually used in object detection? What are the advantages？ Could we use standard cross-entropy with softmax？

That makes sense. In fact, I tried `softmax_focal_loss` following `sigmoid_focal_loss` as follows: ```python def softmax_focal_loss(inputs, targets, num_boxes, alpha: float = 0.25, gamma: float = 2): """ Args: inputs: A float...

thx, now fixed!

Hi, @lxtGH, thx for sharing. Could you please add our new TPAMI 2023 paper on `Domain-aware Segmentation`: SePiCo: Semantic-Guided Pixel Contrast for Domain Adaptive Semantic Segmentation, the link is https://ieeexplore.ieee.org/document/10018569...