search

donydchen / mvsplat

🌊 [ECCV'24] MVSplat: Efficient 3D Gaussian Splatting from Sparse Multi-View Images
https://donydchen.github.io/mvsplat
Other
498 stars 22 forks source link

Question about Cost volume construction #9

Closed 3Der1999 closed 3 months ago

3Der1999 commented 3 months ago

Hello, I don't understand the inverse depth domain and the specific operations of warp. I can't find explanations for the inverse depth domain online. Can you provide some information about them? Thanks! image

donydchen commented 3 months ago

It means that we sample the candidate depth planes uniformly in the inverse depth (i.e., $1/d$) space. The implementation detail is HERE.

This is a general approach in multi-view stereo. By sampling points in the inverse depth space, when we project these points back to the 2D image spaces, they will distribute more uniformly along the epipolar line. The image below (from this paper) provides a better illustration.

Illustration-of-depth-sampling-and-inverse-depth-sampling-a-Depth-sampling-Depth