Summary
In this chapter, we talked about an approach to fitting deformable mesh models to a point cloud. As we have discussed, obtaining meshes from point clouds is usually a standard step in many 3D computer vision pipelines. The fitting approach in this chapter can be used as a simple baseline approach in practice.
From this deformable mesh fitting approach, we learned how to use PyTorch optimization. We also learned about many loss functions and their PyTorch3D implementations, including Chamfer distances, mesh edge loss, mesh Laplacian smoothing loss, and mesh normal consistency loss.
We learned when these loss functions should be used and for what purposes. We saw several experiments for showing how the loss functions affect the final outcome. You are also encouraged to run your own experiments with different combinations of loss functions and weights.
In the next chapter, we will discuss a very exciting 3D deep learning technique called differentiable rendering. Actually...
