There is an error detection issue with the Avia dataset

engcang / FAST-LIO-SAM-QN

A SLAM implementation combining FAST-LIO2 with pose graph optimization and loop closing based on Quatro and Nano-GICP

Other

213 stars 16 forks source link

There is an error detection issue with the Avia dataset #2

Open hr2894235132 opened 10 months ago

hr2894235132 commented 10 months ago

First of all, thank you very much for such a great job! But I had a problem using it to provide .bag for FAST-LIO-Localization-QN, and it showed a wrong correction on the dataset, which looks like a mismatch in the loopback detection module due to the similarity of the scenes? My dataset is using the dataset provided in FAST_LIO_LOCALIZATION

hr2894235132 commented 10 months ago

To add, I'm also having this issue in FAST-LIO-SAM Public, are there any parameters for avia that need to be adjusted?(A mismatch occurred about 1 minute or so into the video.)

https://github.com/engcang/FAST-LIO-SAM-QN/assets/86639226/36b89f8e-7004-459c-a053-fbd0a6c0a0cb

engcang commented 10 months ago

@hr2894235132 Hi. FAST-LIO2 works perfectly on this dataset and so I think we do not need to correct poses with loop-closings. The environment has lots of similar structures and Livox AVIA's FoV is too narrow. I think sadly Quatro cannot work properly on this dataset.

hr2894235132 commented 10 months ago

@hr2894235132 Hi. FAST-LIO2 works perfectly on this dataset and so I think we do not need to correct poses with loop-closings. The environment has lots of similar structures and Livox AVIA's FoV is too narrow. I think sadly Quatro cannot work properly on this dataset.

Thanks for your reply! Does such a bad match affect my saved .bag files (/keyframe_pcd and /keyframe_pose)? Because I see that pose_corrected_eig was used when saving the .bag file.

hr2894235132 commented 10 months ago

Sorry to bother you again, the main thing is that I have been testing and study FAST-LIO-Localization-QN recently, so I have encountered some problems, please excuse me. In addition, I tested another avia dataset.(https://drive.google.com/file/d/1I3ghnN5o5v4RHTkCgP-6U-5pGG7OMl0D/view?usp=drive_link) and moved slowly around the building, and at the end it seemed to establish a loopback detection constraint, but the trajectory was not corrected, I wonder why? rviz_screenshot_2023_12_21-20_29_52 Here's what I tested with FAST-LIO-SAM, there is a significant correction in the z-axis direction, but there is still an error in the horizontal direction.But I think the method with quatro should perform better.