Research map / 3D vision / Face and shape
HybridMQA: Exploring Geometry-Texture Interactions for Colored Mesh Quality Assessment
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.
Proposes HybridMQA, a hybrid full-reference quality assessment model for colored meshes that fuses GCN-learned 3D surface features, rendered into aligned views, with colored projections via cross-attention.
Problem
Full-reference quality assessment of colored (textured or vertex-color) meshes, where existing model-based methods lack appearance information and projection-based methods lack 3D topology understanding, so geometry-texture interactions are missed.
Approach
Combine a CNN base encoder over normal/vertex/texture maps, a GCN over mesh connectivity, and a differentiable feature-graph rendering that produces 3D feature projections pixel-aligned with colored renderings; then apply multi-scale cross-attention between the two branches to model geometry-texture interactions.
Key results
- HybridMQA reaches SRCC 0.892 / PLCC 0.897 on Nehmé et al., beating the best prior method 3D-PSSIM (0.882 / 0.842).
- On SJTU-TMQA, HybridMQA achieves SRCC 0.887 / PLCC 0.896, a 7.7% PLCC gain over 3D-PSSIM.
- Ablation: 3D-only gives SRCC 0.586 and RGB-only gives 0.820, while the full hybrid reaches 0.892 on Nehmé et al.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use HybridMQA instead of 3D-PSSIM | full-reference colored MQA · Nehmé et al. | PLCC: 0.842 -> 0.897 | Helped |
| Use HybridMQA instead of 3D-PSSIM | full-reference colored MQA · SJTU-TMQA | PLCC: 0.832 -> 0.896 | Helped |
| Use HybridMQA instead of Graphics-LPIPS | full-reference colored MQA · TSMD | PLCC: 0.812 -> 0.919 | Helped |
| Use HybridMQA instead of Graphics-LPIPS | full-reference colored MQA · CMDM | PLCC: 0.925 -> 0.927 | Helped |
| Add textural and graph-based 3D representations to the 3D-only configuration (ablation) | HybridMQA · Nehmé et al. | PLCC: 0.595 -> 0.897 | Helped |
| Add 3D representations to the RGB-only configuration (ablation) | HybridMQA · Nehmé et al. | PLCC: 0.846 -> 0.897 | Helped |
| Add ˆf to the quality representation (ablation) | HybridMQA · Nehmé et al. | PLCC: 0.849 -> 0.897 | Helped |
| Add fm to the quality representation (ablation) | HybridMQA · Nehmé et al. | PLCC: 0.874 -> 0.897 | Helped |
| Add the texture map to the model branch (ablation) | HybridMQA · Nehmé et al. | PLCC: 0.872 -> 0.897 | Helped |
| Add the normal map to the model branch (ablation) | HybridMQA · Nehmé et al. | PLCC: 0.866 -> 0.897 | Helped |
| Add the vertex map to the model branch (ablation) | HybridMQA · Nehmé et al. | PLCC: 0.873 -> 0.897 | Helped |
| Add the Base Encoder (ablation) | HybridMQA · Nehmé et al. | PLCC: 0.851 -> 0.897 | Helped |
Surprising findings
- Excluding the 3D representations (RGB-only) causes a large drop, showing the 3D branch matters even though texture is the strongest single signal.
Limitations
- Uses only six perpendicular viewpoints without weighting their perceptual importance.
- Memory consumption scales with mesh size because the GCN processes the entire mesh graph.
- Only full-reference; no-reference extension left for future work.
Experimental setup
Datasets: Nehmé et al. colored mesh dataset SJTU-TMQA TSMD CMDM
Benchmarks: Nehmé et al. SJTU-TMQA TSMD CMDM
Metrics: SRCC PLCC
Baselines: HD MSDM2 FMPD GeodesicPSIM Fu et al. PSNR SSIM VIF LPIPS Graphics-LPIPS 3D-PSSIM
Builds on: Graphics-LPIPS PyTorch3D ResNet50 Morris et al. GNN update rank loss (Sun et al.)
Closest papers by meaning
- Textured Mesh Saliency: Bridging Geometry and Texture for Human Perception in 3D Graphics (AAAI 2025 · similarity 0.86)
- Hi3DEval: Advancing 3D Generation Evaluation with Hierarchical Validity (NeurIPS 2025 · similarity 0.85)
- Mesh-RFT: Enhancing Mesh Generation via Fine-grained Reinforcement Fine-Tuning (NeurIPS 2025 · similarity 0.85)
- MeshGen: Generating PBR Textured Mesh with Render-Enhanced Auto-Encoder and Generative Data Augmentation (CVPR 2025 · similarity 0.84)
- CraftsMan3D: High-fidelity Mesh Generation with 3D Native Diffusion and Interactive Geometry Refiner (CVPR 2025 · similarity 0.84)
- Unique3D: High-Quality and Efficient 3D Mesh Generation from a Single Image (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.