Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Mesh Compression with Quantized Neural Displacement Fields

Sai Karthikey Pentapati, Gregoire Phillips, Alan C. Bovik

TL;DR

The paper tackles 3D mesh compression by encoding a displacement field on a coarse mesh as an implicit neural representation trained per mesh. It leverages Successive Self-Parameterization (SSP) to establish a bijection between coarse and original surfaces and trains a compact MLP to predict displacements, with pruning, 8-bit quantization, and Huffman coding compressing the INR parameters. The method delivers state-of-the-art geometry preservation across a wide range of bitrates (e.g., 4× to 380×, and up to 65× for the XYZ model), outperforming baselines like NGF and QS-DRC while keeping decoding fast. While encoding remains computationally intensive, the approach offers substantial storage savings and a clear path to extending to per-vertex attributes, with SSP ensuring consistent reconstruction on edge-manifold meshes.

Abstract

Implicit neural representations (INRs) have been successfully used to compress a variety of 3D surface representations such as Signed Distance Functions (SDFs), voxel grids, and also other forms of structured data such as images, videos, and audio. However, these methods have been limited in their application to unstructured data such as 3D meshes and point clouds. This work presents a simple yet effective method that extends the usage of INRs to compress 3D triangle meshes. Our method encodes a displacement field that refines the coarse version of the 3D mesh surface to be compressed using a small neural network. Once trained, the neural network weights occupy much lower memory than the displacement field or the original surface. We show that our method is capable of preserving intricate geometric textures and demonstrates state-of-the-art performance for compression ratios ranging from 4x to 380x.

Mesh Compression with Quantized Neural Displacement Fields

TL;DR

The paper tackles 3D mesh compression by encoding a displacement field on a coarse mesh as an implicit neural representation trained per mesh. It leverages Successive Self-Parameterization (SSP) to establish a bijection between coarse and original surfaces and trains a compact MLP to predict displacements, with pruning, 8-bit quantization, and Huffman coding compressing the INR parameters. The method delivers state-of-the-art geometry preservation across a wide range of bitrates (e.g., 4× to 380×, and up to 65× for the XYZ model), outperforming baselines like NGF and QS-DRC while keeping decoding fast. While encoding remains computationally intensive, the approach offers substantial storage savings and a clear path to extending to per-vertex attributes, with SSP ensuring consistent reconstruction on edge-manifold meshes.

Abstract

Implicit neural representations (INRs) have been successfully used to compress a variety of 3D surface representations such as Signed Distance Functions (SDFs), voxel grids, and also other forms of structured data such as images, videos, and audio. However, these methods have been limited in their application to unstructured data such as 3D meshes and point clouds. This work presents a simple yet effective method that extends the usage of INRs to compress 3D triangle meshes. Our method encodes a displacement field that refines the coarse version of the 3D mesh surface to be compressed using a small neural network. Once trained, the neural network weights occupy much lower memory than the displacement field or the original surface. We show that our method is capable of preserving intricate geometric textures and demonstrates state-of-the-art performance for compression ratios ranging from 4x to 380x.

Paper Structure

This paper contains 14 sections, 2 equations, 10 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Neural Mesh Compression
  4. Prerequisite: Successive Self-Parameterization
  5. Method Overview
  6. Generating Training Data
  7. Neural Mesh Compression Architecture
  8. Compressing The INR Parameters
  9. Experiments & Results
  10. Experiment Setup
  11. Ablation Studies
  12. Runtime and Memory Usage
  13. Failure Cases
  14. Conclusion and Scope for Future Research

Figures (10)

  • Figure 1: Successive self-parameterization. (Figure borrowed from neuralsubdivision)
  • Figure 2: Overview of Neural Mesh Compression
  • Figure 3: One-ring feature accumulation
  • Figure 4: Comparison with Neural Progressive Meshes (NPM). The mesh in the left column (green) is the original, the middle column (blue) is NPM, and the right (gray) is ours. The point-to-mesh errors are scaled $\times10^4$
  • Figure 5: Pruning and quantizing NGF leads to a rough and spiky geometry reconstruction.
  • ...and 5 more figures