search

autonomousvision / sdfstudio

A Unified Framework for Surface Reconstruction
Apache License 2.0
1.96k stars 183 forks source link

fail to converge when apply background model #252

Open blacksino opened 10 months ago

blacksino commented 10 months ago

Hello,I have apply foreground mask on my custom dataset and I would like to disable background model to speed up the training,however,it fails to converge. Here is a single example of my custom dataset :

image

When training with background model(grid),it works fine. How could I resolve this kind of problem.

niujinshuchong commented 10 months ago

Hi, I think it should work if you disable background. Did you use black for the background color?

blacksino commented 10 months ago

Yes, I have set background color as black,no idea why it fails when I disable background model. It seems that the object I want to reconstruct is relatively small compared to the bounding box of the axis-aligned bounding box (AABB) obtained after normalizing the camera pose, resulting in excessive learning pressure on the foreground model. If this is the case, can I further trim the AABB using the sparse point cloud obtained from COLMAP?