3D shape representation with DeepSDF
This small project shows how two 3D shapes (.off mesh files) can be voxelized and sdf values can be assigned to them. Then the DeepSDF model is used to train these two shapes in an implicit neural representation. Doing so can provide a to reconstruct and generate shapes with one variable.
Any two shapes (.off files) can be placed in dataset folder to let the DeepSDF have a transition from one shape to other with one single paramter.
To train the model, run train.
Any samples from ModelNet10 (with .off extension) can be used here.
The config is used to put all the config info on it.
The DeepSDF model is placed into this. You can also see notebook file.
In the following, the training is done two times on two pairs of shapes. The first pair is a toilet-bed case and the second contains two chairs. In each case, the first sample is coded 1 and the second sample coded 2 initially. The following is showing the reconstruction of shapes with different codes. The ground truth are somehow shown when the code is 1 or 2.
DeepSDF is a great tool for implicit 3D shapre representation. However, the transition should be smarter than this to avoid meaningless shapes between these two shapes.