Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Hash Grid Feature Pruning

Yangzhi Ma, Bojun Liu, Jie Li, Li Li, Dong Liu

TL;DR

The paper addresses inefficiency in hash-grid representations used with Gaussian splatting due to sparse 3D splat distributions. It proposes a post-training hash grid feature pruning method that identifies valid vertices from input coordinates and uses a hash function $h(X_V) = ((x_V \pi_x) \oplus (y_V \pi_y)) \bmod T$ to map neighbor vertices to indices, encoding only these features. Coordinate normalization to the hash grid is performed with $X_R = \frac{X - X_{\min}}{X_{\max} - X_{\min}} \cdot R$, enabling pruning without affecting reconstruction. On AVS-VRU CTC with the i3DV platform, the approach yields an average bitrate reduction of about $8\%$ (up to $15.5\%$ on some sequences) while maintaining PSNR, demonstrating practical rate-distortion gains for VR/3D streaming.

Abstract

Hash grids are widely used to learn an implicit neural field for Gaussian splatting, serving either as part of the entropy model or for inter-frame prediction. However, due to the irregular and non-uniform distribution of Gaussian splats in 3D space, numerous sparse regions exist, rendering many features in the hash grid invalid. This leads to redundant storage and transmission overhead. In this work, we propose a hash grid feature pruning method that identifies and prunes invalid features based on the coordinates of the input Gaussian splats, so that only the valid features are encoded. This approach reduces the storage size of the hash grid without compromising model performance, leading to improved rate-distortion performance. Following the Common Test Conditions (CTC) defined by the standardization committee, our method achieves an average bitrate reduction of 8% compared to the baseline approach.

Hash Grid Feature Pruning

TL;DR

The paper addresses inefficiency in hash-grid representations used with Gaussian splatting due to sparse 3D splat distributions. It proposes a post-training hash grid feature pruning method that identifies valid vertices from input coordinates and uses a hash function to map neighbor vertices to indices, encoding only these features. Coordinate normalization to the hash grid is performed with , enabling pruning without affecting reconstruction. On AVS-VRU CTC with the i3DV platform, the approach yields an average bitrate reduction of about (up to on some sequences) while maintaining PSNR, demonstrating practical rate-distortion gains for VR/3D streaming.

Abstract

Hash grids are widely used to learn an implicit neural field for Gaussian splatting, serving either as part of the entropy model or for inter-frame prediction. However, due to the irregular and non-uniform distribution of Gaussian splats in 3D space, numerous sparse regions exist, rendering many features in the hash grid invalid. This leads to redundant storage and transmission overhead. In this work, we propose a hash grid feature pruning method that identifies and prunes invalid features based on the coordinates of the input Gaussian splats, so that only the valid features are encoded. This approach reduces the storage size of the hash grid without compromising model performance, leading to improved rate-distortion performance. Following the Common Test Conditions (CTC) defined by the standardization committee, our method achieves an average bitrate reduction of 8% compared to the baseline approach.
Paper Structure (6 sections, 3 equations, 4 figures, 1 table)

This paper contains 6 sections, 3 equations, 4 figures, 1 table.

Table of Contents

  1. Introduction
  2. Method
  3. Experiments
  4. Experimental Conditions
  5. Performance Evaluation
  6. Conclusion

Figures (4)

  • Figure 1: The valid and invalid vertices in hash grid. The hash grid leverages the features of vertices surrounding each Gaussian splat to compute the feature at that location via linear interpolation. Due to the non-uniform distribution of Gaussian splats, a large number of vertices in the grid do not participate in computation, and their corresponding features are thus designated as invalid.
  • Figure 2: Hash grid adoption in i3DV. (a) depicts the use of a hash grid to assist entropy modeling for static Gaussian splat compression. (b) illustrates the employment of a hash grid to represent an implicit residual field for dynamic Gaussian splat compression.
  • Figure 3: Combining pruning strategy with compression. Based on the input positions, invalid features are identified and pruned, leaving only the valid features to be encoded and decoded.
  • Figure 4: Rate-distortion Comparison. Rate-Distortion (R-D) curves of different methods. 250 frames are evaluated across a wide range of size following the CTC AVS-VRU-CTC.