search

princeton-vl / DROID-SLAM

BSD 3-Clause "New" or "Revised" License
1.66k stars 273 forks source link

Perform not very well using custom data source. #61

Open buenos-dan opened 1 year ago

buenos-dan commented 1 year ago

Thanks for open source. I want to have a try in my custom data, but I'm stuck now. I record a video using realsense D455 in a building then I config everything rightly(remove image distortion, calc camera intrinsics etc.). But the trajectory has a big difference with ground truth. So I wonder if anybody encounter the same problem? And If anybody has got the nice performance using the custom data source, please comment and let me know, I'm very appreciate for that. PS: I try with realsense D455 and fisheye camera. here are some exp results.

left-image is tested with EuRoc dataset(MH_03), is as perfect as paper said. middle-image is tested with our fisheye camera, the shape is around, but is not smoothing. right-image is tested with realsense D455, the shape is very different with groundtruth. The custom data recorded in normal building sense. I wonder how to improve the performance without train the model again.

lilly5791 commented 1 year ago

Please check this issue that I wrote. https://github.com/princeton-vl/DROID-SLAM/issues/63#issue-1321712482

My custom data works well on demo visualization. However, the trajectory result is very different with the ground truth when I visualize with my own matlab code.