search

dcharatan / pixelsplat

[CVPR 2024 Oral, Best Paper Runner-Up] Code for "pixelSplat: 3D Gaussian Splats from Image Pairs for Scalable Generalizable 3D Reconstruction" by David Charatan, Sizhe Lester Li, Andrea Tagliasacchi, and Vincent Sitzmann
http://davidcharatan.com/pixelsplat/
MIT License
864 stars 60 forks source link

How flexible is Min Baseline Requirement #48

Closed skrya closed 7 months ago

skrya commented 7 months ago

Hello David,

Thanks for the work! I see that the minimum baseline requirement is 1e-3 on the datasets. May I know if the method can still produce reasonable result if the baseline is say > 1e-5 but less than 1e-3, which would happen in an autonomous driving setting?

Thanks!

dcharatan commented 7 months ago

It might still work if you try it out. The main concern is that the code that computes the epipolar lines could become numerically unstable with very small baselines. If I'm remembering correctly, I added baseline_epsilon to the RE10k dataset because I figured any baseline smaller than that would correspond to basically no movement anyway (given the scale distribution in RE10k), meaning that smaller baselines wouldn't provide a useful training signal.