Closed bin70 closed 4 years ago
Hi, Jinbin,
This is about the receptive field of USIP. The definition of keypoint
depends on the scale you are looking at. For example, the center of a
scene/object is actually a keypoint if we are looking at a large scale. But
locally, the center is not a meaningful keypoint. As mentioned in USIP,
this phenomenon is called degeneracy
. That is, usually we are looking for
local keypoints, but USIP will degenerate into global keypoints (like
center, princial axis) if the receptive field is too large.
The way to prevent such degeneracy is to limit the receptive field. More specifically, small M and/or large K will lead to large receptive field, or vice versa.
In your case, your testing data is different from the training KITTI data. Therefore you have to tune the M, K at testing phrase, so that you are getting reasonable results. You may consider increasing M and/or reducing K in testing.
Thanks.
Best Regards Jiaxin Li
Tan Jinbin notifications@github.com 于2019年11月10日周日 下午2:22写道:
Hello, Jiaxin: I trained USIP detector on kitti dataset and tested it on some lidar frames collected in a tunnle scence. But there is a large offset between the estimated keypoints and the origin cloud (I can be sure that they are in a uniform coordinate system.) :
[image: 1] https://user-images.githubusercontent.com/27135261/68539641-649fff00-03c1-11ea-9cb7-4092c004ef97.jpg
[image: 2] https://user-images.githubusercontent.com/27135261/68539642-6cf83a00-03c1-11ea-94d9-444b3970f536.jpg
I konw you mentioned in paper that it's not necessary for all keypoints to be any of the points in original point cloud. But I'm a bit confused that the distance between them shoudn't be so large under the constraint of the Point-to-Point loss described in paper.
How do I set up the network parameters to avoid so large the distance between them? Do I need to create a training set for the tunnel scene by myself?
Can you give me some hints? Thanks a lot!
— You are receiving this because you are subscribed to this thread. Reply to this email directly, view it on GitHub https://github.com/lijx10/USIP/issues/9?email_source=notifications&email_token=AC5G72LKYJX4DCTZ6C4MOELQS6SDDA5CNFSM4JLLDK3KYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HYHD6IA, or unsubscribe https://github.com/notifications/unsubscribe-auth/AC5G72MAF4KAEZU5GIFPWMTQS6SDDANCNFSM4JLLDK3A .
-- 黎嘉信 Jiaxin LI PhD, National University of Singapore B.S., Tsinghua University Mobile: 65-9094 1909 / 86-15201519053 Email: lijx1992@gmail.com / jli@u.nus.edu jli@nus.edu.sg
Thanks for your patient reply. I changed (M,K) according to your suggestion, but the offset is still too large:
I ended up solving this problem by training on my own dataset:
Thanks agains!
Hello, Jiaxin: I trained USIP detector on kitti dataset and tested it on some lidar frames collected in a tunnle scence. But there is a large offset between the estimated keypoints and the origin cloud (I can be sure that they are in a uniform coordinate system.) :
I konw you mentioned in paper that it's not necessary for all keypoints to be any of the points in original point cloud. But I'm a bit confused that the distance between them shoudn't be so large under the constraint of the Point-to-Point loss described in paper.
How do I set up the network parameters to avoid so large the distance between them? Do I need to create a training set for the tunnel scene by myself?
Can you give me some hints? Thanks a lot!