Research map / 3D vision / Face and shape
Learning 3D Shape Fidelity Metric from Real-world Distortions
Xuelu Feng, Tianyu Luan, Zixin Zhu, Akshobhya Sharma, Phani Nuney, Junsong Yuan, Chunming Qiao
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add LoRA-style finetuning of the pretrained PointNet MLPs (ablation) | LoCaSE · RSF main (train/test split) | PLCC: 0.671 -> 0.728 | Improved |
| add local attention within vertex mesh neighborhoods (ablation) | LoCaSE · RSF main (train/test split) | PLCC: 0.605 -> 0.728 | Improved |
| use both local attention and LoRA (ablation) | LoCaSE · RSF main (train/test split) | PLCC: 0.597 -> 0.728 | Improved |
| use mesh-topology local attention instead of ball query (ablation) | LoCaSE · RSF main (train/test split) | PLCC: 0.645 -> 0.728 | Improved |
| train on real-distortion RSF data instead of synthetic ShapeGrading only (ablation) | LoCaSE · RSF main (train/test split) | PLCC: 0.380 -> 0.728 | Improved |
| use ModelNet-pretrained PointNet plus LoRA instead of training from scratch (ablation) | LoCaSE · RSF main (train/test split) | PLCC: 0.697 -> 0.728 | Improved |
| replace simple PointNet backbone with MeshCNN, DiffusionNet, PointNet++ or DGCNN (ablation) | LoCaSE · RSF main (train/test split) | PLCC: 0.728 -> 0.576 (MeshCNN), 0.541 (DiffusionNet), 0.592 (PointNet++), 0.699 (DGCNN) | Hurt |
| set LoRA rank to 32 instead of 8 or 16 (ablation) | LoCaSE · RSF main (train/test split) | SROCC: 0.7256 -> 0.7573 (rank 8 vs 32) | Improved |
| increase LoRA rank from 32 to 64 (ablation) | LoCaSE · RSF main (train/test split) | SROCC: 0.7573 -> 0.7274 | No clear effect |
| weight Smooth L1 loss 5 versus PLCC and SROCC losses 1 each (ablation) | LoCaSE · RSF main (train/test split) | PLCC: 0.677 -> 0.728 | Improved |
| use LoCaSE instead of Chamfer Distance on unseen objects and methods | LoCaSE · RSF test-only | PLCC: -0.46 -> 0.68 | Improved |
Closest papers by meaning
- Textured Geometry Evaluation: Perceptual 3D Textured Shape Metric via 3D Latent-Geometry Network (AAAI 2026 · similarity 0.90)
- SRAM: Shape-Realism Alignment Metric for No Reference 3D Shape Evaluation (AAAI 2026 · similarity 0.89)
- Evaluating Multiview Object Consistency in Humans and Image Models (NeurIPS 2024 · similarity 0.86)
- Front2Back: Single View 3D Shape Reconstruction via Front to Back Prediction (CVPR 2020 · similarity 0.86)
- Flow-Corrected Shape Optimization: Taming Manifold Drift in High-Dimensional 3D Models (NeurIPS 2026 · similarity 0.85)
- Multistable Shape from Shading Emerges from Patch Diffusion (NeurIPS 2024 · similarity 0.85)
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.
