About overlap

[wuchunxu123]

Hi, thanks for your excellent work，it's really impressive. I have a question. It is mentioned that you select image pairs having overlap from 40% to 80% for training. Is the overlap value used to balance the numbers of matched points and unmatched points? I noticed that the pos part and neg part share the same weight in the loss function. If it is, can I use the matching percentage of keypoints in Scannet to represent the overlap and generate image pairs？I think this is more efficient and can directly balances the number of matched and unmatched points. Looking forward to your reply.

[sarlinpe]

On ScanNet the overlap is unrelated to the ratio of matchable and unmatchable keypoints. As referenced in a previous reply, equally balancing the positive and negative losses can help the training.