Logo
ScienceStack
Table of Contents
Fetching ...
Sign In

From Far and Near: Perceptual Evaluation of Crowd Representations Across Levels of Detail

Xiaohan Sun, Carol O'Sullivan

TL;DR

This study addresses how humans perceive crowd representations rendered at different levels of detail and distances. By combining a controlled user study with quantitative metrics (PSNR, SSIM, LPIPS) and a memory-footprint analysis, it compares Mesh, Impostor, NeRF, and 3D Gaussian representations across LoD levels. The results show neural representations (NeRF, 3DGS) closely approximate a high-resolution mesh at high detail, while image-based impostors excel in efficiency at distance; 3D Gaussians offer a robust mid-to-high LoD option, and traditional meshes excel at near-view fidelity. The work delivers a perception-driven LoD pipeline and toolkit, guiding designers to balance fidelity and performance in real-time crowd rendering and enabling adaptive, perceptually informed LoD scheduling.

Abstract

In this paper, we investigate how users perceive the visual quality of crowd character representations at different levels of detail (LoD) and viewing distances. Each representation: geometric meshes, image-based impostors, Neural Radiance Fields (NeRFs), and 3D Gaussians, exhibits distinct trade-offs between visual fidelity and computational performance. Our qualitative and quantitative results provide insights to guide the design of perceptually optimized LoD strategies for crowd rendering.

From Far and Near: Perceptual Evaluation of Crowd Representations Across Levels of Detail

TL;DR

This study addresses how humans perceive crowd representations rendered at different levels of detail and distances. By combining a controlled user study with quantitative metrics (PSNR, SSIM, LPIPS) and a memory-footprint analysis, it compares Mesh, Impostor, NeRF, and 3D Gaussian representations across LoD levels. The results show neural representations (NeRF, 3DGS) closely approximate a high-resolution mesh at high detail, while image-based impostors excel in efficiency at distance; 3D Gaussians offer a robust mid-to-high LoD option, and traditional meshes excel at near-view fidelity. The work delivers a perception-driven LoD pipeline and toolkit, guiding designers to balance fidelity and performance in real-time crowd rendering and enabling adaptive, perceptually informed LoD scheduling.

Abstract

In this paper, we investigate how users perceive the visual quality of crowd character representations at different levels of detail (LoD) and viewing distances. Each representation: geometric meshes, image-based impostors, Neural Radiance Fields (NeRFs), and 3D Gaussians, exhibits distinct trade-offs between visual fidelity and computational performance. Our qualitative and quantitative results provide insights to guide the design of perceptually optimized LoD strategies for crowd rendering.
Paper Structure (17 sections, 5 figures, 7 tables)

This paper contains 17 sections, 5 figures, 7 tables.

Table of Contents

  1. Introduction
  2. Related Work
  3. Data Representations for Crowd Rendering
  4. Level of Detail (LoD) for Animated Crowds
  5. Implementation Methods
  6. Mesh
  7. Impostor
  8. NeRF
  9. Gaussian Splatting
  10. Experimental Design
  11. User Study
  12. Quantitative Experiments
  13. Visual Comparison
  14. Summary.
  15. Memory Usage
  16. ...and 2 more sections

Figures (5)

  • Figure 1: Comparison of four crowd rendering representations (geometric meshes, image-based impostors, Neural Radiance Fields, and 3D Gaussians) across four levels of detail (LoD 0-3). LoD 0 corresponds to the highest resolution (100 %), with detail halved at each subsequent level (50 %, 25 %, 12.5 %). The examples illustrate how fidelity, structural clarity, and smoothness degrade under simplification, revealing characteristic differences among representation types.
  • Figure 2: Examples of data.
  • Figure 3: Examples of mesh LoD.
  • Figure 4: Example of video stimulus.
  • Figure 5: Video block: mean proportion chosen for each representation across LoD and Distance (D0--D4). Error bars show $\pm$ 1 SE across participants.