This repository contains source code of my master thesis: Learning 3D Shape Completion under Self-Supervision. The goal of this project is to complete 3D shape from partial input(i.e., 3D point cloud). We build a generative framework based on encoder-decoder network to learn a shared latent shape space that can be used to generate complete and realistic 3D shapes from partial and sparse inputs.
The datasets that are used:
To configure the parameters:
cd configs/default.yaml
To train the model:
python train.py
To generate meshes from the trained model:
python generate.py
To evaluate the trained model:
python eval.py
im2mesh
: evaluation tool from occupancy networks