Research map / Label-efficient and robust learning / Recognition and long-tail
Manifold Embedding of Deep Image Features for Image Matching via Neural Adjoint Maps
Nafie El Amrani, Moritz Heep, Florian Bernard
Read the paperCheck your idea against 137,217 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 a manifold-based embedding of deep image features, using neural adjoint maps, to regularise semantic image correspondence beyond plain nearest-neighbour matching on foundation model features.
Problem
Semantic correspondence between different object instances of similar categories in real-world images is hard, and nearest-neighbour matching on foundation model features is insufficiently regularised.
Approach
Embeds deep image features on a manifold and uses neural adjoint maps to regularise matching, rather than relying on raw nearest-neighbour matching in feature space. Only the abstract is available, so details are limited.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use proposed manifold-embedding LBO instead of the image graph Laplacian (ablation) | NAMI with NAM · SPair-71k | PCK@0.1 (%): 38.96 -> 56.50 | Improved |
| use neural adjoint map instead of linear SmoothFM (ablation) | NAMI with proposed LBO · SPair-71k | PCK@0.1 (%): 50.13 -> 56.50 | Improved |
| use NAMI instead of Cheng et al. | Cheng et al. functional maps for images · TSS | runtime (s): 95.6 -> 0.76 | Improved |
| use NAMI instead of Cheng et al. at 50 eigenfunctions | Cheng et al. · TSS | PCK@0.05 (%): 60.3 -> 84.3 | Improved |
| increase eigenfunctions beyond 50 (ablation) | NAMI · TSS | PCK@0.05 (%): 84.3 (k=50) -> 80.0 (k=200) | Hurt |
| add Neural ZoomOut refinement (ablation) | NAMI · TSS | PCK@0.05 (%): 84.3 -> 86.3 at roughly 10x runtime | Improved |
| add feature refinement network from Cheng et al. (ablation) | NAMI · TSS | PCK@0.05 (%): 86.72 -> 81.84 | Hurt |
| use NAMI versus best baseline on large rotations | NAMI · Faust-2D hard split | PCK | No clear effect |
| use NAMI instead of Fuse2Match | Fuse2Match · SPair-71k | PCK@0.1 (%): 61.0 -> 56.5 | Hurt |
Closest papers by meaning
- Deep Geometric Functional Maps: Robust Feature Learning for Shape Correspondence (CVPR 2020 · similarity 0.90)
- Deep Functional Maps: Structured Prediction for Dense Shape Correspondence (ICCV 2017 · similarity 0.89)
- Memory-Scalable and Simplified Functional Map Learning (CVPR 2024 · similarity 0.87)
- Hybrid Functional Maps for Crease-Aware Non-Isometric Shape Matching (CVPR 2024 · similarity 0.87)
- Denoising Functional Maps: Diffusion Models for Shape Correspondence (CVPR 2025 · similarity 0.87)
- Spectral Meets Spatial: Harmonising 3D Shape Matching and Interpolation (CVPR 2024 · similarity 0.87)
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.
