Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

Robust Mesh Saliency GT Acquisition in VR via View Cone Sampling and Geometric Smoothing

Guoquan Zheng, Jie Hao, Huiyu Duan, Yongming Han, Liang Yuan, Dong Zhang, Guangtao Zhai

TL;DR

This work tackles the challenge of obtaining robust 3D mesh saliency ground truth in VR by addressing two key limitations of prior methods: (1) aliasing and discontinuities from single-ray sampling, and (2) topological leakage during smoothing. It introduces View Cone Sampling (VCS) to emulate the foveal receptive field with a Gaussian ray bundle and a Hybrid Manifold-Euclidean Constrained Diffusion (HCD) to propagate gaze influence along mesh geodesics while respecting topology. The approach yields a topology-aware saliency field by combining cumulative gaze density with geodesic diffusion and a face-vertex smoothing pipeline, validated on 100 textured meshes across varying resolutions, with significant gains in sAUC, CC, KL, and IC over baselines. The results offer a high-fidelity, VR-aligned GT baseline for 3D mesh saliency research and practical implications for perceptual coding, mesh simplification, and VR rendering, with public release of data and code to foster reproducibility.

Abstract

Reliable 3D mesh saliency ground truth (GT) is essential for human-centric visual modeling in virtual reality (VR). However, current 3D mesh saliency GT acquisition methods are generally consistent with 2D image methods, ignoring the differences between 3D geometry topology and 2D image array. Current VR eye-tracking pipelines rely on single ray sampling and Euclidean smoothing, triggering texture attention and signal leakage across gaps. This paper proposes a robust framework to address these limitations. We first introduce a view cone sampling (VCS) strategy, which simulates the human foveal receptive field via Gaussian-distributed ray bundles to improve sampling robustness for complex topologies. Furthermore, a hybrid Manifold-Euclidean constrained diffusion (HCD) algorithm is developed, fusing manifold geodesic constraints with Euclidean scales to ensure topologically-consistent saliency propagation. By mitigating "topological short-circuits" and aliasing, our framework provides a high-fidelity 3D attention acquisition paradigm that aligns with natural human perception, offering a more accurate and robust baseline for 3D mesh saliency research.

Robust Mesh Saliency GT Acquisition in VR via View Cone Sampling and Geometric Smoothing

TL;DR

This work tackles the challenge of obtaining robust 3D mesh saliency ground truth in VR by addressing two key limitations of prior methods: (1) aliasing and discontinuities from single-ray sampling, and (2) topological leakage during smoothing. It introduces View Cone Sampling (VCS) to emulate the foveal receptive field with a Gaussian ray bundle and a Hybrid Manifold-Euclidean Constrained Diffusion (HCD) to propagate gaze influence along mesh geodesics while respecting topology. The approach yields a topology-aware saliency field by combining cumulative gaze density with geodesic diffusion and a face-vertex smoothing pipeline, validated on 100 textured meshes across varying resolutions, with significant gains in sAUC, CC, KL, and IC over baselines. The results offer a high-fidelity, VR-aligned GT baseline for 3D mesh saliency research and practical implications for perceptual coding, mesh simplification, and VR rendering, with public release of data and code to foster reproducibility.

Abstract

Reliable 3D mesh saliency ground truth (GT) is essential for human-centric visual modeling in virtual reality (VR). However, current 3D mesh saliency GT acquisition methods are generally consistent with 2D image methods, ignoring the differences between 3D geometry topology and 2D image array. Current VR eye-tracking pipelines rely on single ray sampling and Euclidean smoothing, triggering texture attention and signal leakage across gaps. This paper proposes a robust framework to address these limitations. We first introduce a view cone sampling (VCS) strategy, which simulates the human foveal receptive field via Gaussian-distributed ray bundles to improve sampling robustness for complex topologies. Furthermore, a hybrid Manifold-Euclidean constrained diffusion (HCD) algorithm is developed, fusing manifold geodesic constraints with Euclidean scales to ensure topologically-consistent saliency propagation. By mitigating "topological short-circuits" and aliasing, our framework provides a high-fidelity 3D attention acquisition paradigm that aligns with natural human perception, offering a more accurate and robust baseline for 3D mesh saliency research.
Paper Structure (13 sections, 7 equations, 4 figures, 2 tables)

This paper contains 13 sections, 7 equations, 4 figures, 2 tables.

Table of Contents

  1. Introduction
  2. Eye-tracking Data Acquisition
  3. Eye-tracking Data Acquisition Environment
  4. View Cone Sampling Strategy
  5. Mesh Saliency GT Modeling
  6. Geometric Preprocessing and Normalization
  7. Gaze Density Measurement
  8. Geodesic Gaussian Diffusion
  9. Face-Vertex Dual Smoothing
  10. Experimental Results and Discussion
  11. Comparative Analysis of Qualitative Results
  12. Comparative Analysis of Quantitative Results
  13. CONCLUSION

Figures (4)

  • Figure 1: (a) Discrepancy between perceptual mechanism and single ray sampling method. (b) Sparse geometric structure may introduce significant discontinuous saliency or penetrating accidentally ray collision. (c) Ignoring the obstacle of geometric gaps on visual attention.
  • Figure 2: (a) Example of the Unity3D eye-tracking data acquisition scene. (b) Schematic cross section of the VCS strategy. (c) Example of ray distribution within the sampling field of the VCS strategy. (d) Pipeline for 3D mesh saliency GT generation.
  • Figure 3: (a) Qualitative comparison visualization of saliency maps on representative 3D mesh models. (b) Comparison of capture methods (Post-processing as HCD). (c) Comparison of post-processing methods (The data is sampled via VCS).
  • Figure 4: Statistical comparison of sampling coverage efficacy. (a) Face counts ranging from 5k to 50k. (b) Face counts ranging from 400k to 800k.