Research map / Vision-language and multimodal / Human and embodied modeling
Mesh Mamba: A Unified State Space Model for Saliency Prediction in Non-Textured and Textured Meshes
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 a VR eye-tracking mesh saliency dataset with paired non-textured and textured conditions and proposes Mesh Mamba, a bidirectional SSM with subgraph embedding for face-level saliency prediction.
Problem
Mesh saliency prediction has mostly targeted non-textured meshes, and there is little data or method for how texture and geometry jointly shape visual attention on 3D meshes.
Approach
A unified SSM (Mamba) architecture operating on mesh-face tokens: subgraph embedding via farthest point sampling plus random walk sampling preserves topology, a bidirectional SSM with feature diffusion/aggregation models global context, and UV-based texture alignment fuses texture with geometric features.
Key results
- On the textured mesh benchmark with texture features, Mesh Mamba reaches CC 0.6305 and SIM 0.7232, versus 0.5605 CC for the next best (MeshNet).
- On non-textured geometry-only saliency prediction, Mesh Mamba achieves CC 0.6140 vs 0.5993 for Mamba3D.
- On SAL3D, Mesh Mamba achieves CC 0.6763 vs 0.6513 for MeshNet++.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add texture features | Mesh Mamba · Textured mesh saliency | CC: 0.6066 -> 0.6305 | Helped |
| Add vertex color features | Mesh Mamba · Non-textured mesh saliency | CC: 0.6140 -> 0.5868 | Hurt |
| Add texture features | Mesh Mamba · Non-textured mesh saliency | CC: 0.6140 -> 0.6080 | No clear effect |
| Add vertex color features | Mesh Mamba · Textured mesh saliency | CC: 0.6066 -> 0.5957 | Hurt |
| Add texture features | Mesh Mamba · Non-textured mesh saliency | CC: 0.6140 -> 0.6080 | Hurt |
| Add shape features (ablation) | Mesh Mamba · Textured mesh saliency | CC: 0.5403 -> 0.6305 | Helped |
| Add spatial features (ablation) | Mesh Mamba · Textured mesh saliency | CC: 0.6175 -> 0.6305 | Helped |
| Add curve features (ablation) | Mesh Mamba · Textured mesh saliency | CC: 0.6036 -> 0.6305 | Helped |
| Use the Mamba backbone instead of a Transformer (ablation) | Mesh Mamba · Textured mesh saliency | CC: 0.6113 -> 0.6305 | Helped |
| Add the graph convolution layer (ablation) | Mesh Mamba · Textured mesh saliency | CC: 0.5993 -> 0.6305 | Helped |
| Add subgraph embedding (ablation) | Mesh Mamba · Textured mesh saliency | CC: 0.6237 -> 0.6305 | Helped |
| Add feature diffusion and aggregation (ablation) | Mesh Mamba · Textured mesh saliency | CC: 0.5889 -> 0.6305 | Helped |
Surprising findings
- Adding vertex colors or texture features degraded performance for most methods on non-textured meshes.
- Vertex colors gave limited benefit even on textured meshes, and texture alignment did not improve Mesh Mamba on non-textured meshes.
Limitations
- Texture alignment did not help the model on non-textured meshes, and vertex colors offered limited gains.
- Dataset collected from 60 non-computer-graphics participants on Free3D assets; scale and generalization beyond these assets are not discussed in depth.
Experimental setup
Datasets: Free3D meshes (custom saliency dataset) SAL3D
Benchmarks: custom textured/non-textured mesh saliency dataset SAL3D
Metrics: CC SIM KLD SE (MSE)
Baselines: PointNet PointNet2-SSG PointNet2-MSG PointTrans PointMixer StraTrans MeshNet MeshNet++ DiffusionNet-xyz DiffusionNet-hks DSM SAL3D
Builds on: Mamba MeshNet DiffusionNet Mamba3D
Closest papers by meaning
- Textured Mesh Saliency: Bridging Geometry and Texture for Human Perception in 3D Graphics (AAAI 2025 · similarity 0.89)
- Samba: A Unified Mamba-based Framework for General Salient Object Detection (CVPR 2025 · similarity 0.86)
- Spatial-Mamba: Effective Visual State Space Models via Structure-Aware State Fusion (ICLR 2025 · similarity 0.85)
- Cue3D: Quantifying the Role of Image Cues in Single-Image 3D Generation (NeurIPS 2025 · similarity 0.85)
- MPOD123: One Image to 3D Content Generation Using Mask-enhanced Progressive Outline-to-Detail Optimization (CVPR 2024 · similarity 0.84)
- Demystify Mamba in Vision: A Linear Attention Perspective (NeurIPS 2024 · similarity 0.84)
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.