lijx10 / PCLKeypoints

Python wrapper for some point cloud keypoint detectors of PCL, including ISS, Harris3D, SIFT3D
MIT License
87 stars 20 forks source link

segmentation fault #1

Open XuyangBai opened 5 years ago

XuyangBai commented 5 years ago

Hi @lijx10 Thank you so much for your sharing. I tried PCLKeypoint and installed it successly, but I always get segmentation fault whenever I tried to use keypointSift or other functions. I have tried on Ubuntu and Mac but neither works (for Mac, I got segmentation fault when I import it). I wonder have you ever met such problems?

屏幕快照 2019-10-29 上午11 29 36

屏幕快照 2019-10-29 上午11 29 00
lijx10 commented 5 years ago

Hi, have you solved the segfault? I haven't encounter such problem before. One possible problem is whether the point cloud is Nx3 or 3xN? I am not sure.

XuyangBai notifications@github.com 于 2019年10月29日周二 下午12:30写道:

Hi @lijx10 https://github.com/lijx10 Thank you so much for your sharing. I tried PCLKeypoint and installed it successly, but I always get segmentation fault whenever I tried to use keypointSift or other functions. I have tried on Ubuntu and Mac but neither works (for Mac, I got segmentation fault when I import it). I wonder have you ever met such problems?

[image: 屏幕快照 2019-10-29 上午11 29 36] https://user-images.githubusercontent.com/15166089/67735536-68399a80-fa3f-11e9-9ed3-398082b9cd00.png [image: 屏幕快照 2019-10-29 上午11 29 00] https://user-images.githubusercontent.com/15166089/67735538-696ac780-fa3f-11e9-819d-0ae07642e8c2.png

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/lijx10/PCLKeypoints/issues/1?email_source=notifications&email_token=AC5G72JPEFC33JGFWQYTFJLQQ6U67A5CNFSM4JGC6Z2KYY3PNVWWK3TUL52HS4DFUVEXG43VMWVGG33NNVSW45C7NFSM4HU6ARJQ, or unsubscribe https://github.com/notifications/unsubscribe-auth/AC5G72M4TQNAAOCZQ3IOWH3QQ6U67ANCNFSM4JGC6Z2A .

XuyangBai commented 4 years ago

Hi @lijx10 It is a weird problem: I can't import both open3d and PCLKeypoint in the same file. Maybe it is a bug of open3d. When I do not use open3d and PCLKeypoint at the same time, everything works well. BTW, may I ask what's the parameter(like min_scale, min_contrast) you used for hand crafted methods(iss, sift, harris, sift) when evaluating repeatability on redwood dataset? I think the parameter in save_keypoints.py is for outdoor scenes.

Thank you so much.

lijx10 commented 4 years ago

This is a known issue of incompatibility between open3d and pytorch. You have to import open3d before importing pytorch, otherwise there will be segfault. Probably this is related to the cuda initialization done by both open3d and pytorch.

Configurations of redwood: Basically we do a param search for all methods to generate the best possible repeatability for them.

ISS: number of keypoints | parameters | repeatability at matching radius r=0.05, average-min-max | repeatability at matching radius r=0.1, average-min-max 4 0.6/0.6 0.0347-0-1 0.078-0-1 6 0.5/0.5 0.0294-0-1 0.069-0-1 9 0.4/0.4 0.0424-0-0.6667 0.092-0-1 17 0.3/0.3 0.0406-0-0.4545 0.1-0-0.5 41 0.2/0.2 0.0574-0-0.3333 0.142-0-0.5 77 0.15/0.15 0.0655-0-0.3148 0.19-0.02-0.56 125 0.12/0.12 0.0765-0-0.3065 0.23-0.03-0.581 180 0.1/0.1 0.097/0.074/0.2785 0.292-0.08-0.63 209 0.05/0.1 0.0883-0-0.3089 0.285-0.02-0.73 291 0.1/0.05 0.1137-0-0.488 0.278-0-0.82 434 0.05/0.05 0.1522-0-0.513 0.375-0.03-0.94 Harris3D 6 0.4/0.001 0.055-0-1 0.1152-0-1 8 0.3/0.001 0.1094-0-1 0.202-0-1 13 0.2/0.001 0.1314-0-0.75 0.2264-0-1 20 0.15/0.001 0.1453-0-0.8 0.2528-0-0.9286 45 0.1/0.001 0.1473-0-1 0.2713-0-1 131 0.05/0.001 0.0966-0-0.513 0.2301-0-0.8113 257 0.05/0.00001 0.1071-0-0.4508 0.2704-0.0124-0.7872 478 0.05/1e-8 0.0912-0.0017-0.3784 0.2568-0.0336-0.7387

SIFT3D 5 0.05-4-8-0.03 0.0046-0-1 0.016-0-1 7 0.05-4-8-0.025 0.0092-0-0.4 0.025-0-0.5 14 0.05-4-8-0.015 0.0253-0-0.4444 0.07-0-0.67 25 0.05-4-8-0.01 0.0353-0-0.4706 0.086-0-0.57 41 0.05-4-8-0.0075 0.0389-0-0.2692 0.105-0-0.54 87 0.05-4-8-0.005 0.0599-0-0.3889 0.155-0-0.6 128 0.05-4-8-0.004 0.0705-0-0.3333 0.191-0-0.5 238 0.05-4-8-0.0025 0.0899-0-0.315 0.263-0-0.6 468 0.05-4-8-1e-8 0.1253-0.0229-0.4114 0.364-0.09-0.81

XuyangBai notifications@github.com 于2019年11月1日周五 上午1:33写道:

Hi @lijx10 https://github.com/lijx10 It is a weird problem: I can't import both open3d and PCLKeypoint in the same file. Maybe it is a bug of open3d. When I do not use open3d and PCLKeypoint at the same time, everything works well. BTW, may I ask what's the parameter(like min_scale, min_contrast) you used for hand crafted methods(iss, sift, harris, sift) when evaluating repeatability on redwood dataset? I think the parameter in save_keypoints.py is for outdoor scenes.

Thank you so much.

— You are receiving this because you were mentioned. Reply to this email directly, view it on GitHub https://github.com/lijx10/PCLKeypoints/issues/1?email_source=notifications&email_token=AC5G72ITU4E67GBWNORSPYLQRMCHHA5CNFSM4JGC6Z2KYY3PNVWWK3TUL52HS4DFVREXG43VMVBW63LNMVXHJKTDN5WW2ZLOORPWSZGOECYNRGY#issuecomment-548460699, or unsubscribe https://github.com/notifications/unsubscribe-auth/AC5G72LEY7K62IP5PK4GBIDQRMCHHANCNFSM4JGC6Z2A .

-- 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

XuyangBai commented 4 years ago

@lijx10 Thank you so much for your help !

XuyangBai commented 4 years ago

Hi @lijx10 Could you please also share the configurations of KITTI or you just use the same parameter defined in save_keypoints .py for all number of keypoints situations ?