Colmap script

[sonnefred]

Hi, do you use convert.py you provided to get camera poses? When I use this script, the result is very sparse, only few images are kept. Could you give any hint on this? Thanks.

[kcheng1021]

Hi, what dataset or scene do you use? It your scene has sparse observation or your images are hard to extract keypoints, then your result will be very sparse.

[sonnefred]

Hi, what dataset or scene do you use? It your scene has sparse observation or your images are hard to extract keypoints, then your result will be very sparse.

Hi, I used scene data of city streets like Waymo, but it's collected by ourselves. How did you get the colmap result which you provided? When I use the convert.py, the result is very sparse. Thanks.

[kcheng1021]

I just run the convert.py with the default parameters. I think there may be some reasons for the sparsity:

1. Your car's speed is high, thus leading to the sparser captures of a scene, compared to Waymo's low speed car
2. The keypoints are harder to get for your scene, due to the textureless plane, low quality of images, small resolution of images, and so on... You can make sure the resolution of your image is higher enough (like more than 1080p), and replace the SIFT keypoints with superpoint in COLMAP for better keypoint extraction (like using hloc[https://github.com/cvg/Hierarchical-Localization]).
3. Change some parameters in convert.py. For this part, you could refer to colmap for more infos. convert.py indeed make some simplification for colmap's default setting, one simple way to check is that using colmap's automatic reconstruction choice and check if it works.