Research map / Label-efficient and robust learning / Zero-/few-shot and meta-learning
Zero-Shot Image Feature Consensus with Deep Functional Maps
Xinle Cheng, Congyue Deng, Adam Harley, Yixin Zhu, Leonidas Guibas
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 diagonal compactness and consistency regularizers to descriptor loss (ablation) | Feature functional map · SPair-71k | PCK@0.1: 44.6 -> 55.3 | Improved |
| add both regularizers on top of one regularizer (ablation) | Feature functional map · SPair-71k | MSE: 97.9 -> 88.0 | Improved |
| add only a single regularizer (diag or cons) over no regularization (ablation) | Feature functional map · SPair-71k | MSE: 95.3 -> 97.9 / 100.3 | Hurt |
| use different feature sets for basis (SD) and loss (DINOv2) instead of the same set (ablation) | Feature functional map · TSS | PCK avg: 71.0 -> 78.9 | Improved |
| choose SD for basis and DINOv2 for loss rather than reverse (ablation) | Feature functional map · TSS | PCK avg: 72.2 -> 78.9 | Improved |
| use functional map with a single feature set versus raw nearest neighbors (ablation) | DINOv2-ViT-B/14 · TSS | PCK avg: 69.4 -> 71.0 | Improved |
| fuse features via functional map instead of concatenation | DINOv2-ViT-B/14 + SD · TSS | smoothness: 6.81 -> 5.74 | Improved |
| fuse DINOv2-ViT-S/14 and SD via functional map instead of concatenation | DINOv2-ViT-S/14 + SD · TSS | PCK@0.05: 78.1 -> 71.5 | Hurt |
| fuse different layers of the same network via functional map instead of concatenation | DINOv2-ViT-B/14 · TSS | EPE: 30.9 -> 25.9 | Improved |
| functional map versus concatenation on sparse keypoints | DINOv2-ViT-B/14 + SD · SPair-71k | PCK@0.1: 57.2 -> 55.3 | Mixed |
Closest papers by meaning
- Weakly-Supervised Learning of Dense Functional Correspondences (ICCV 2025 · similarity 0.89)
- Deep Functional Maps: Structured Prediction for Dense Shape Correspondence (ICCV 2017 · similarity 0.89)
- Manifold Embedding of Deep Image Features for Image Matching via Neural Adjoint Maps (NeurIPS 2026 · similarity 0.88)
- The Functional Correspondence Problem (ICCV 2021 · similarity 0.88)
- Memory-Scalable and Simplified Functional Map Learning (CVPR 2024 · similarity 0.87)
- SemAlign3D: Semantic Correspondence between RGB-Images through Aligning 3D Object-Class Representations (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.
