Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Neural Pruning for 3D Scene Reconstruction: Efficient NeRF Acceleration

Tianqi Ding, Dawei Xiang, Pablo Rivas, Liang Dong

TL;DR

NeRFs deliver high-fidelity 3D reconstructions but require long training times. This work systematically evaluates pruning-based compression for NeRF, comparing uniform sampling, importance pruning, and coreset-based methods to reduce MLP size and accelerate training. The key finding is that coreset-driven pruning achieves about $50\%$ reduction in model size and about $35\%$ faster training with only a small drop in PSNR. These results show pruning as a practical tool to enable NeRF deployment in resource-constrained settings and motivate broader compression strategies for 3D reconstruction.

Abstract

Neural Radiance Fields (NeRF) have become a popular 3D reconstruction approach in recent years. While they produce high-quality results, they also demand lengthy training times, often spanning days. This paper studies neural pruning as a strategy to address these concerns. We compare pruning approaches, including uniform sampling, importance-based methods, and coreset-based techniques, to reduce the model size and speed up training. Our findings show that coreset-driven pruning can achieve a 50% reduction in model size and a 35% speedup in training, with only a slight decrease in accuracy. These results suggest that pruning can be an effective method for improving the efficiency of NeRF models in resource-limited settings.

Neural Pruning for 3D Scene Reconstruction: Efficient NeRF Acceleration

TL;DR

NeRFs deliver high-fidelity 3D reconstructions but require long training times. This work systematically evaluates pruning-based compression for NeRF, comparing uniform sampling, importance pruning, and coreset-based methods to reduce MLP size and accelerate training. The key finding is that coreset-driven pruning achieves about reduction in model size and about faster training with only a small drop in PSNR. These results show pruning as a practical tool to enable NeRF deployment in resource-constrained settings and motivate broader compression strategies for 3D reconstruction.

Abstract

Neural Radiance Fields (NeRF) have become a popular 3D reconstruction approach in recent years. While they produce high-quality results, they also demand lengthy training times, often spanning days. This paper studies neural pruning as a strategy to address these concerns. We compare pruning approaches, including uniform sampling, importance-based methods, and coreset-based techniques, to reduce the model size and speed up training. Our findings show that coreset-driven pruning can achieve a 50% reduction in model size and a 35% speedup in training, with only a slight decrease in accuracy. These results suggest that pruning can be an effective method for improving the efficiency of NeRF models in resource-limited settings.

Paper Structure

This paper contains 21 sections, 4 equations, 4 figures, 4 tables, 1 algorithm.

Table of Contents

  1. Introduction
  2. Related work
  3. Neural Network Pruning
  4. Unstructured Pruning
  5. Structured Pruning
  6. Pruning Methods
  7. Uniform Sampling
  8. Importance Pruning
  9. Coreset
  10. Neural Radiance Fields (NeRF)
  11. Model compression on NeRF
  12. Methods and Experiments
  13. Sparcity in NeRF MLP
  14. Pruning the edges
  15. Problem of Unstructured Edge Pruning
  16. ...and 6 more sections

Figures (4)

  • Figure 1: A visualization of the MLP architecture from mildenhall2021nerf. The network accepts $\gamma(x)$, a positional encoding of the 3D coordinates, along with $\gamma(d)$, an encoding of the viewing direction. It then outputs the density $\sigma$ and RGB color of the point $(x,y,z)$.
  • Figure 2: Frequency Histogram of edge weights in NeRF. A large percent of edge weights fall into the range $[0, 0.05]$.
  • Figure 3: An illustration of the importance weight calculation. Red arrows represent incoming edges for neurons in layer $i$, whereas green arrows indicate outgoing edges.
  • Figure 4: Performance comparison visualization for different compression scales.