NMSFreeCoder of DETR3D: cls_scores.sigmoid() then cls_scores.view(-1).topk(max_num)

[Ricardo-Yu]

Hi, I'm confuesd about the NMSFreeCoder of DETR3D, which use cls_scores.sigmoid() then cls_scores.view(-1).topk(max_num) to generate final prediction bbox. Then when select topK, how to avoid won't select the same query multiple times? e.g. if cls_scores.shape=[300, 10] and max_num=100, then after sigmoid(), it's possible that an object_query have multiple high cls_scores since it's using sigmoid instead of softmax. After .view(-1) to flat as [3000], when apply topK, it's possible to select an object_query twice if it has 2 high cls scores from sigmoid?

If an object_query may selected twice from bbox coder, then the DETR3D will predict 2 bbox3d fully overlapped, but with different classes? Will this be an issue when using in autonomous driving tasks, which may need mutually exclusive classification for downstream to react accordingly?

mmdetection3d/projects/DETR3D/detr3d/nms_free_coder.py

Lines 55 to 56 in fe25f7a

	cls_scores = cls_scores.sigmoid()
	scores, indexes = cls_scores.view(-1).topk(max_num)