search

oppo-us-research / SpacetimeGaussians

[CVPR 2024] Spacetime Gaussian Feature Splatting for Real-Time Dynamic View Synthesis
https://oppo-us-research.github.io/SpacetimeGaussians-website/
Other
508 stars 40 forks source link

Pose conversion on custom dataset #22

Open adkAurora opened 4 months ago

adkAurora commented 4 months ago

I'd like to consult about a camera pose conversion issue while conducting experiments on the dna dataset. I've rewritten the dna dataset into colmap format for training, but the results are poor. Here are the poses: image

lizhan17 commented 4 months ago

we have plan to release comstom dataset, please stay tuned for a custom dataset loader and preprocessing step.

The pose convention is a colmap convention. I can tell some important factors that may affect the quality during the exps:

1) camera model types. n3d is undistorted centered, techni is undistorted uncentered. immersive is distorted uncentered.

2) the per-frame consistency of same poses and points. colmap's poses/world coordinates (for the same camera) is not consistent across different runs even you use fixed external option. the only thing you can do is feeding the camera poses to colmap and just use point_triangulator to get sparse points (see what we did in n3d..). such that can make sure that same static point should be at same position across different time.