acherednychenko / mosaic-sdf

Unofficial implementation of Mosaic-SDF algorithm

MosaicSDF Implementation

Introduction

Repository contains implementation of Mosaic-SDF shape representation algorhithm, as described in Mosaic-SDF for 3D Generative Models paper.

Mosaic-SDF (M-SDF): a simple 3D shape representation that approximates the Signed Distance Function (SDF) of a given shape by using a set of local grids spread near the shape's boundary. The M-SDF representation is fast to compute for each shape individually making it readily parallelizable; it is parameter efficient as it only covers the space around the shape's boundary; and it has a simple matrix form, compatible with Transformer-based architectures.

What you will find in sources:

• MosaicSDF algorhithm to approximate SDF using local grids: mosaic_sdf.py
• Optimization of grids parameters - position and scale: optimizer.py
  • Gives marginal improvements, as random sampling of grids already gives decent results
• Parameters fine-tuning using Ray Tune: run_tune.py
• Interactive use of the algorhithm: interactive_mosaic_sdf.ipynb
• Helper tools:
  • Sampling SDFs
  • Visualizing SDFs (using marching-cubes) and meshes

What is not implemented:

• Farthest point sampling
• Building generative model on top of SDFs

Getting Started

Pre-requisites

• Install PyTorch 3D
• Install graphviz,

# On linux, run:
sudo apt-get install graphviz

• then run pip install -r requirements.txt to install required packages

Results

Comparison of MosaicSDF(mosaic_meshes) with 256 mosaic cells to ground-truth sdf(gt_sdf_mesh) and ground-truth mesh (gt_mesh):