search

WU-CVGL / BAD-Gaussians

[ECCV 2024] "BAD-Gaussians: Bundle Adjusted Deblur Gaussian Splatting". ⚡Train a scene from real-world blurry images in minutes!
https://lingzhezhao.github.io/BAD-Gaussians/
Apache License 2.0
137 stars 4 forks source link

Using hloc for camera pose #10

Open KevinOPG opened 2 months ago

KevinOPG commented 2 months ago

Hello, thank you for the amazing works. This is not an issue, but just wondering have you tried used the hloc toolbox for ns-process-data instead of colmap. If you did, was it better than colmap for your blurred datasets?

LingzheZhao commented 2 months ago

Hi, we haven't tested it yet, but I believe hloc with more robust learning-based feature detectors/matchers will give better pose estimation initializations for BAD-Gaussians. With those extremely blurred images, I expect hloc less likely to fail. (For example, in our challenging demo dataset ikun, colmap failed on 105 of 306 images, maybe I will try with this dataset later.)

KevinOPG commented 2 months ago

Hi, we haven't tested it yet, but I believe hloc with more robust learning-based feature detectors/matchers will give better pose estimation initializations for BAD-Gaussians. With those extremely blurred images, I expect hloc less likely to fail. (For example, in our challenging demo dataset ikun, colmap failed on 105 of 306 images, maybe I will try with this dataset later.)

Thanks! I will try it when I get a chance as well. 怎么哪里都有长劲鹿啊

1van2ha0 commented 1 month ago

Hello! Could you please tell me more about some robust sfm algorithms, just like hloc toolbox? I have trouble dealing with my dataset. I cannot get camera poses by colmap. So I'm trying to find other ways to get camera poses. Thanks!

LingzheZhao commented 1 month ago

ns-process-data in nerfstudio has support for hloc + superpoint & lightglue, etc, but their documentation doesn't cover these yet. FYI, some examples can be found in the related PR and issues (https://github.com/nerfstudio-project/nerfstudio/pull/1026 and https://github.com/nerfstudio-project/nerfstudio/issues/1884).

Also, ns-process-data currently does not support pre-calibrated camera intrinsics, while we found it important for COLMAP given challenging inputs (few shot or motion blur). In our previous experiments, we follow COLMAP's documentation (here) and fix the intrinsics manually. Maybe I can try adding this to ns-process-data and add more details to our README in the future.