Research map / 3D vision / Face and shape
Textured Mesh Saliency: Bridging Geometry and Texture for Human Perception in 3D Graphics
Read the paperCheck your idea against 108,967 papers
AI-extracted summary. Written by AI from the paper's full text. Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Builds the first textured-mesh saliency dataset from a 6-DOF VR eye-tracking study (100 models, 30 participants) and proposes a face-level network that fuses UV-aligned texture features with geometric features to predict per-face saliency.
Problem
Mesh saliency prediction has focused on non-textured (or vertex-colored) meshes, so human visual attention on richly textured 3D meshes is unmodeled and no dataset or predictor exists for it.
Approach
Treat each triangular face as a unit; align texture patches to faces via UV-based implicit interpolation over a latent code map, extract spatial, shape and multi-ring normal-based structural geometry features, and fuse them with cross-attention and neighbor aggregation to predict per-face saliency density.
Key results
- On the texture stage, the proposed model reaches CC 0.5764, SIM 0.7061, KLD 0.3198 and SE 0.0273, outperforming the best baseline MeshNet (CC 0.5605).
- Geometry-only results: proposed model CC 0.5675 vs MeshNet 0.5512 and PointMixer 0.5261.
- Removing shape features drops CC from 0.5675 to 0.4135, the largest ablation drop.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add texture features (ablation) | Ours · Textured mesh saliency test set | CC: 0.5675 -> 0.5764 | Helped |
| add spatial features (ablation) | Ours · Textured mesh saliency test set | CC: 0.5682 -> 0.5764 | Helped |
| add shape features (ablation) | Ours · Textured mesh saliency test set | CC: 0.4135 -> 0.5764 | Helped |
| add structural features (ablation) | Ours · Textured mesh saliency test set | CC: 0.5625 -> 0.5764 | Helped |
| use the aggregation module instead of direct concatenation (ablation) | Ours · Textured mesh saliency test set | CC: 0.5585 -> 0.5764 | Helped |
| add a latent code and use it instead of sampled raw texture images (ablation) | Ours texture alignment module · Textured mesh saliency test set | CC: 0.5530 -> 0.5764 | Helped |
| expand the sampling grid (ablation) | Ours texture alignment module · Textured mesh saliency test set | CC: 0.5637 -> 0.5764 | Helped |
| use implicit interpolation instead of nearest-neighbor sampling (ablation) | Ours texture alignment module · Textured mesh saliency test set | CC: 0.5708 -> 0.5764 | Helped |
Surprising findings
- Adding vertex colors decreased performance for the proposed model (CC 0.5675 to 0.5571), and most baselines also declined with color input.
- Point-based methods did not benefit from texture features at all, while face-based methods did.
Limitations
- Small dataset of 100 models with 80/20 split limits generality.
- Vertex color features did not help, suggesting limited feature richness for textured content.
- Participants were newcomers to computer graphics, possibly limiting domain generalization.
Experimental setup
Datasets: Free3d TurboSquid
Benchmarks: textured mesh saliency dataset (self-built)
Metrics: CC SIM KLD SE
Baselines: PointNet PointNet2-SSG PointNet2-MSG PointTrans PointMixer StraTrans MeshNet MeshNet++ DiffusionNet-xyz DiffusionNet-hks DSM SAL3D
Builds on: DSM SAL3D MeshNet DiffusionNet PointNet++
Closest papers by meaning
- Mesh Mamba: A Unified State Space Model for Saliency Prediction in Non-Textured and Textured Meshes (CVPR 2025 · similarity 0.89)
- MAtCha Gaussians: Atlas of Charts for High-Quality Geometry and Photorealism From Sparse Views (CVPR 2025 · similarity 0.86)
- HybridMQA: Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment (CVPR 2025 · similarity 0.86)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.86)
- DreamCraft3D: Hierarchical 3D Generation with Bootstrapped Diffusion Prior (ICLR 2024 · similarity 0.86)
- TriTex: Learning Texture from a Single Mesh via Triplane Semantic Features (CVPR 2025 · similarity 0.86)
Writing a paper in this area? scime draft is a free LaTeX editor with AI included, real-time co-authors and a check of your references and venue rules before you submit. It also imports OpenAI Prism projects.