The algorithm is not working properly on my dataset

[Lousin]

Hello, I have consulted you by email before, and thank you very much for your reply.
I have successfully run the EuRoC example, but when I use my own data set, the map has been unable to initialize, I tried to solve the problem from the following ways, but all failed.

1, I modified orbmatcher.cpp in the code based on the issue in the orbslam (raulmur/ORB_SLAM2#59) .
2, my image contrast is low, so I tried to lower the ORB Extractor: Fast threshold, but it still didn't work. (Two adjacent images can only match 10 feature points)
Here are the images I used, how can I improve the algorithm to solve my problem?

[patriksc]

Hey @Lousin, usually, there should not be a need to change the parameters of the ORB extractor.The content in you image seems a bit noisy IMO, buy there should probably be enough structure to detect ORB features. What I realize, downloading one of your images, it shows me a resolution of 1920x1080, while the content itself is like 500x400, which means only approx. 10% of your image contain content, while the rest is black. This is certainly not ideal, and I'd recommend to make sure most of the image contains "matchable" structure and not black background. That's probably the most important thing to do. Then, I'd do some visual debugging - print intermediate steps, e.g. images with the extracted keypoints shown, to verify things look OK. Finally, make sure to hav enough motion in your dataset to allow the algorithm to initialize.