search

nerlfield / wild-gaussian-splatting

DUSt3R + Gaussian Splatting
185 stars 18 forks source link

Generalizability of data #1

Open Mr-ZhuJun opened 5 months ago

Mr-ZhuJun commented 5 months ago

Brother, I would like to ask, you should have tried different data. Is the current Dust3r+3DGS method stable on different data?

nerlfield commented 5 months ago

Hi @Mr-ZhuJun! I've tested this approach across various scenes. I've tried 3D reconstruction of human heads, and it worked well, although there was some jittering in the final result. I believe that this might be due to Dust3r generating cameras with varying intrinsic parameters.

Therefore, to enhance stability, there's a significant opportunity to adjust Dust3r by optimizing a single set of intrinsic parameters for all cameras.

Mr-ZhuJun commented 5 months ago

I have tested Dust3r using fixed internal parameters, but the accuracy of external parameter estimation has not improved. Currently, I plan to use some recent work of 3DGS to provide feedback and adjust the pose.

Bin-ze commented 2 months ago

Hi, I also tried to experiment on dust3, with more than 20 personal collection scenes, both small and large.

But my experimental results show that dust3r is far from colmap's results in any scene, when colmap estimation does not have obvious errors.

I would like to know what improvements you have made on this basis to make dust3r work on your scene？