sashatr/sdf_playground

Research Signed Distance Functions

SDF with NN

---

The task is to simulate the SFD function for transforming the mesh 3D data format into the SFD.

---

Input data (mesh)

Chair	Cube

Pyramid	Teddy

---

I used fully-connected and convolutional neural networks to model SDF functions for input data.

Models:

• Net7 - NN with with seven fully-connected layers
• Net12 - NN with with twelve fully-connected layers
• ConvNet - NN with with three convolutional layers
• ConvNet2 - NN with with six convolutional layers

Below comparison is the result of training four models for four input mesh figures.

Cube Mesh	Pyramid Mesh	Teddy Mesh	Chair Mesh

SDF heat map for Cube	SDF heat map for Pyramid	SDF heat map for Teddy	SDF heat map for Chair

Net7 output for Cube	Net7 output for Pyramid	Net7 output for Teddy	Net7 output for Chair

Net12 output for Cube	Net12 output for Pyramid	Net12 output for Teddy	Net12 output for Chair

CNN output for Cube	CNN output for Pyramid	CNN output for Teddy	CNN output for Chair

CNN2 output for Cube	CNN2 output for Pyramid	CNN2 output for Teddy	CNN2 output for Chair

---

Conclusions:

• Even the simplest fully-connected models with a small number of layers are able to simulate SFD functions.
• One-dimensional convolution is an effective method for this task. Even without normalization, it describes the mesh well, and the weight of the model is less than the mesh file.
• After a large number of experiments, such models will describe mesh objects well using less resources.
• This is a great area for research ;)

---

TODO:

• Improve loss func (e.g. from DeepSDF)
• Improve visualization
• Parallelize inference models
• More convolution network experiments
• Add nn.BatchNorm1d
• Use a model for group mesh objects
• Etc

---