Research map / Learning theory and algorithms / Prediction and decision losses
Repeatability Is Not Enough: Learning Affine Regions via Discriminability
Dmytro Mishkin, Filip Radenovic, Jiri Matas
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 |
|---|---|---|---|
| replace hardest-negative gradient (HardNeg) with HardNegC (ablation) | AffNet, HardNet descriptor · HSequences hardest pairs | correct matches: n/c -> 388 | Improved |
| use HardNegC instead of PosDist loss for affine shape (ablation) | AffNet, SIFT descriptor · HSequences hardest pairs | correct matches: n/c -> 386 | Improved |
| learn affine shape with AffNet instead of Baumberg | Hessian + HardNet · HSequences hardest pairs | correct matches: 298 -> 388 | Improved |
| learn orientation instead of dominant gradient orientation | Hessian + HardNet · HSequences hardest pairs | correct matches: 339 -> 387 | Improved |
| predict residual shape A'' with zero-bias parametrization instead of full matrix A (ablation) | AffNet · HPatchesSeq | correct matches: 360 -> 388 | Improved |
| learn orientation separately after shape instead of jointly (ablation) | AffNet with A' and R(alpha) · HPatchesSeq | correct matches: 250 -> 327 | Improved |
| learn orientation instead of dominant gradient (ablation) | AffNet A'' parameterization · HPatchesSeq | correct matches: 349 -> 388 | Improved |
| use AffNet instead of Baumberg | DoG detector · HPatches viewpoint | repeatability: 0.51 -> 0.52 | Improved |
| use AffNet instead of Baumberg | Hessian detector · HPatches viewpoint | repeatability: 0.52 -> 0.56 | Improved |
| replace Baumberg with AffNet | HardNet++ BoW retrieval · Oxford5k BoW | mAP: 60.8 -> 68.3 | Improved |
| replace Baumberg with AffNet | TFeat-M* BoW retrieval · Oxford5k BoW | mAP: 46.7 -> 45.5 | Mixed |
| keep all AffNet features instead of restricting to Baumberg count (ablation) | HesAffNet-HardNet++ · Oxford5k BoW | mAP: 64.3 -> 68.3 | Improved |
Closest papers by meaning
- Learning Affine Correspondences by Integrating Geometric Constraints (CVPR 2025 · similarity 0.87)
- Making Affine Correspondences Work in Camera Geometry Computation (ECCV 2020 · similarity 0.87)
- AffIn-Space: Learning Affine-Invariant Representations for 3D Spatial Understanding with MLLMs (ICML 2026 · similarity 0.86)
- PUMP: Pyramidal and Uniqueness Matching Priors for Unsupervised Learning of Local Descriptors (CVPR 2022 · similarity 0.86)
- Detect-To-Retrieve: Efficient Regional Aggregation for Image Search (CVPR 2019 · similarity 0.85)
- Learning Discriminative and Transformation Covariant Local Feature Detectors (CVPR 2017 · 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.
