Point Cloud Fusion from depth maps

In one of the issues described ([here](https://github.com/bmild/nerf/issues/35)), you have been successful in fusing point clouds from depth images. You have mentioned that you have used "MVSNet techniques to filter noisy points" and find there is another repo "kwea123/CasMVSNet_pl" which seems to do that but "eval.py" requires a trained MVSNet.

Could you please explain your methodology of rendering a point cloud using a set of depth maps and images (without any additional re-training) for point cloud fusion ? Could you also please update/publish the related codebase ?

Start with the linked issue #35 and the referenced eval.py in kwea123/CasMVSNet_pl to understand the existing MVSNet-based filtering context. Define how the depth maps and images should be fused without retraining, then publish the related codebase and document the methodology and usage; the issue does not name files in nerf_pl.

Written by the indexing model from the issue text.