Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Plug-and-Play Acceleration of Occupancy Grid-based NeRF Rendering using VDB Grid and Hierarchical Ray Traversal

Yoshio Kato, Shuhei Tarashima

TL;DR

This work tackles the computational burden of rendering NeRFs by replacing dense occupancy grids with OpenVDB-based sparse grids and employing Hierarchical DDA for efficient ray traversal, enabling plug-and-play acceleration without re-training. By pruning redundancies and prioritizing finer voxels within multi-resolution VDB grids, the method achieves notable FPS gains on both NeRF-Synthetic (~12% average) and Mip-NeRF 360 (~4–5% average) while preserving image quality. Key contributions include leveraging VDB’s hierarchical structure for OG representation and integrating HDDA traversal with two ray-sampling kernels, highlighting that VDB+HDDA can outperform the baseline in many scenarios. The approach offers practical speedups for trained OG-based NeRF models with modest memory overhead, suitable for deployment on GPUs, though benefits can vary with scene size and overhead from multi-grid setups.

Abstract

Transmittance estimators such as Occupancy Grid (OG) can accelerate the training and rendering of Neural Radiance Field (NeRF) by predicting important samples that contributes much to the generated image. However, OG manages occupied regions in the form of the dense binary grid, in which there are many blocks with the same values that cause redundant examination of voxels' emptiness in ray-tracing. In our work, we introduce two techniques to improve the efficiency of ray-tracing in trained OG without fine-tuning. First, we replace the dense grids with VDB grids to reduce the spatial redundancy. Second, we use hierarchical digital differential analyzer (HDDA) to efficiently trace voxels in the VDB grids. Our experiments on NeRF-Synthetic and Mip-NeRF 360 datasets show that our proposed method successfully accelerates rendering NeRF-Synthetic dataset by 12% in average and Mip-NeRF 360 dataset by 4% in average, compared to a fast implementation of OG, NerfAcc, without losing the quality of rendered images.

Plug-and-Play Acceleration of Occupancy Grid-based NeRF Rendering using VDB Grid and Hierarchical Ray Traversal

TL;DR

This work tackles the computational burden of rendering NeRFs by replacing dense occupancy grids with OpenVDB-based sparse grids and employing Hierarchical DDA for efficient ray traversal, enabling plug-and-play acceleration without re-training. By pruning redundancies and prioritizing finer voxels within multi-resolution VDB grids, the method achieves notable FPS gains on both NeRF-Synthetic (~12% average) and Mip-NeRF 360 (~4–5% average) while preserving image quality. Key contributions include leveraging VDB’s hierarchical structure for OG representation and integrating HDDA traversal with two ray-sampling kernels, highlighting that VDB+HDDA can outperform the baseline in many scenarios. The approach offers practical speedups for trained OG-based NeRF models with modest memory overhead, suitable for deployment on GPUs, though benefits can vary with scene size and overhead from multi-grid setups.

Abstract

Transmittance estimators such as Occupancy Grid (OG) can accelerate the training and rendering of Neural Radiance Field (NeRF) by predicting important samples that contributes much to the generated image. However, OG manages occupied regions in the form of the dense binary grid, in which there are many blocks with the same values that cause redundant examination of voxels' emptiness in ray-tracing. In our work, we introduce two techniques to improve the efficiency of ray-tracing in trained OG without fine-tuning. First, we replace the dense grids with VDB grids to reduce the spatial redundancy. Second, we use hierarchical digital differential analyzer (HDDA) to efficiently trace voxels in the VDB grids. Our experiments on NeRF-Synthetic and Mip-NeRF 360 datasets show that our proposed method successfully accelerates rendering NeRF-Synthetic dataset by 12% in average and Mip-NeRF 360 dataset by 4% in average, compared to a fast implementation of OG, NerfAcc, without losing the quality of rendered images.
Paper Structure (11 sections, 1 figure, 3 tables, 2 algorithms)

This paper contains 11 sections, 1 figure, 3 tables, 2 algorithms.

Table of Contents

  1. Introduction
  2. Related Works
  3. Efficient Ray-Sampler
  4. Sparse Data Structures
  5. Enumerating Line-Grid Intersections
  6. Method
  7. Experimental Results
  8. Experimental Setup
  9. Results
  10. Discussion
  11. Variants of Ray-Sampling Kernel

Figures (1)

  • Figure 1: Examples of ray-marching algorithms. DDA (left), Chessboard Distance (CD)cd (center), and HDDA on quadtree (right). Black dots depict lookup operations. Black lines are where the colors are sampled. White voxel means empty space. In DDA, we need lookup the voxel's occupancy one by one. CD saves the $L_\infty$ distance to the nearest occupied voxel in advance. HDDA can skip by power of two according to the information in the quadtree.