Network degeneration question

[Leo-820]

I wonder is the network able to use different M and K value during training and testing? And if I understand correctly, the value M equal to the number of keypoint you got. So why USIP doesnot degenerate when detect only 4 keypoints (as experiments in Figure 4), since you mentioned that small value for M will increase the receptive field and cause the FPN to degenerate. Is it because you choose a proper value of K to avoid degeneration?

[lijx10]

1. The network is able to use different M and/or K during training and testing. Specifically, the output will be similar as long as the receptive field is roughly the same. The network would still work even when the receptive field is significantly different, like twice larger or smaller. Of course there will be unexpected output when the difference is too extreme.
2. We apply an additional constraint, i.e., keypoint-on-pointcloud loss, to reduce the degenerate effect. That's why we can still get proper keypoints with M=4.