-
Chamfer distance
-
Sampling for visualization
-
Marching cube for visualization
- We use the mesh data structure provided by trimesh for processing the mesh data. To install trimesh, please install trimesh as well as its dependencies using the following two cmds.
conda install -c conda-forge trimesh
and
conda install -c conda-forge scikit-image shapely rtree pyembree
-
There are some mesh models for your test. All models were retrieved from common-3d-test-models. Please follow the licenses when you are using the models.
-
Use
train.py --cfg config_3d.jsonto train the model for 3d shapes
File dataset.py was added to show you how you can write a customized dataset.
File network.py was added to show how a 8-layer MLP can be made. (It has no dropout or other weight initialization/normalization)
File geometry.py was added to re-organize the codes; geometry related functions and classes are moved to this package/file
File basics.py was modified to use the newly added files.
FIle basics.py is a piece of code that you can play with. It is adapted from this repo: https://github.com/Oktosha/DeepSDF-explained
Note that "basics.py" is not the re-implementation of DeepSDF (https://github.com/facebookresearch/DeepSDF) but more similar to the paper (https://arxiv.org/abs/2009.09808) in the sense that the network is trained to overfit a single shape (but 2D in our case).