Open Chatoyant19 opened 4 years ago
hey, I am not sure about your problem because your issue is a very common situation, many reasons cause the initialization method to fail. The stereo or RGBD dataset doesn't need to be initialized, so the stereo dataset is easy to run.
One possible reason is the few numbers of matched keypoints, maybe you can change the keypoints and descriptors type to increase the matching performance.
Another possible cause is in the SVD decomposition part, ORB-SLAM often fails to find the best solution from decomposition, you can output the value in the initialization part, see what things cause the failure, then you can add the improvement to fix the problem.
@gxytcrc Thank you for the quick reply. In the initialization part, I output some logs,and found that the numbers of matched keypoints are sufficient. The main reason for the initialization failure is the reconstruct homography and fundamental. Indeed, as you said, ORB-SLAM fails to find the best solution. I tried to lower the qualification to initialize successfully, but it did not work. I don't know if I should make some changes
Hi, When you run orb-slam3 with yourself monocular datasets, have you encountered the problem of not being able to initialize? with the same bag file, stereo orb-slam can run successfully. I can ensure that the monocular intrinsics are configured correctly. I tried to change my image encoding format from BAYER_RGGB8 to BGR8 and other format, it can only make a slight improvement. But most of the time it's not useful.