Research map / Label-efficient and robust learning / Recognition and long-tail
Online Invariance Selection for Local Feature Descriptors
Rémi Pautrat, Viktor Larsson, Martin R. Oswald, Marc Pollefeys
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 soft assignment of meta descriptors instead of hard assignment (ablation) | LISRD · HPatches | HEstimation: 0.762 -> 0.784 | Improved |
| add meta descriptors instead of greedy selection of local descriptors (ablation) | LISRD · HPatches | HEstimation: 0.774 -> 0.784 | Improved |
| add online selection over 4 descriptors instead of best single descriptor (ablation) | LISRD · HPatches | HEstimation: 0.778 -> 0.784 | Improved |
| add 3x3 tiling to meta descriptors (ablation) | LISRD · HPatches | HEstimation: 0.752 -> 0.784 | Improved |
| use 3x3 tiles instead of 5x5 tiles (ablation) | LISRD · HPatches | HEstimation: 0.773 -> 0.784 | Improved |
| disentangle four invariance descriptors instead of a single invariant descriptor (ablation) | LISRD · HPatches | HEstimation: 0.766 -> 0.784 | Improved |
| apply invariance selection to SIFT and Upright SIFT | SIFT · HPatches viewpoint | precision | Improved |
| compare LISRD with SuperPoint on illumination | SuperPoint · HPatches illumination | recall: 0.565 -> 0.655 | Improved |
| compare LISRD with SuperPoint on viewpoint | SuperPoint · HPatches viewpoint | precision: 0.595 -> 0.626 | Improved |
| LISRD vs SOSNet on homography estimation (coarse meta tiles) | SOSNet · HPatches viewpoint | HEstimation: 0.742 -> 0.688 | Hurt |
| LISRD vs SuperPoint on day reference | SuperPoint · DNIM day ref | recall: 0.178 -> 0.317 | Improved |
| use LISRD descriptors on SIFT keypoints | Up-Root SIFT · Aachen Day-Night | % localized at 0.5m, 2deg: 54.1 -> 72.4 | Improved |
Earlier methods this paper compares with
Named baselines from earlier accepted papers in this paper's results tables: better than 2 of 2 on most shared results.
- SOSNet — SOSNet: Second Order Similarity Regularization for Local Descriptor Learning (CVPR 2019) · better on 4 of 6 shared results
- HardNet — HARD-Net: Hardness-AwaRe Discrimination Network for 3D Early Activity Prediction (ECCV 2020) · better on 4 of 4 shared results
Matched by method name in results tables read by AI. Authors' own tables tend to favour the newer method, and a comparison can be missed when a method is renamed.
Closest papers by meaning
- Comparative Evaluation of Hand-Crafted and Learned Local Features (CVPR 2017 · similarity 0.86)
- Local Descriptors Optimized for Average Precision (CVPR 2018 · similarity 0.86)
- Beyond Cartesian Representations for Local Descriptors (ICCV 2019 · similarity 0.86)
- Learning Discriminative and Transformation Covariant Local Feature Detectors (CVPR 2017 · similarity 0.85)
- Selective Disk Bispectrum: A Complete and Rotation Invariant Image Descriptor (NeurIPS 2026 · similarity 0.84)
- Focus on Local: Finding Reliable Discriminative Regions for Visual Place Recognition (AAAI 2025 · similarity 0.83)
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.
