Research map / 3D vision / Face and shape
GeoDesc: Learning Local Descriptors by Integrating Geometry Constraints
Zixin Luo, Tianwei Shen, Lei Zhou, Siyu Zhu, Runze Zhang, Yao Yao, Tian Fang, Long Quan
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 |
|---|---|---|---|
| train with crowd-sourced SfM data (ablation) | HardNet framework · HPatches | matching mAP: 52.8 -> 57.0 | Improved |
| train with crowd-sourced SfM data (ablation) | HardNet framework · HPatches | verification mAP: 88.4 -> 90.1 | Improved |
| add proposed batch construction with hardest-in-batch loss (ablation) | Config 2 · HPatches | matching mAP: 57.0 -> 50.2 | Hurt |
| add proposed batch construction with hardest-in-batch loss (ablation) | Config 2 · HPatches | retrieval mAP: 73.2 -> 70.4 | Hurt |
| add modified structured loss E1 to batch construction (ablation) | Config 3 · HPatches | matching mAP: 50.2 -> 58.5 | Improved |
| add modified structured loss E1 to batch construction (ablation) | Config 3 · HPatches | verification mAP: 89.9 -> 90.9 | Improved |
| add geometric loss E2 (ablation) | Config 4 · HPatches | matching mAP: 58.5 -> 59.1 | Improved |
| add geometric loss E2 (ablation) | Config 4 · HPatches | retrieval mAP: 74.5 -> 74.9 | Improved |
| replace distance margin with distance ratio in loss (ablation) | structured loss · triplet validation accuracy | training efficiency | Improved |
| use GeoDesc instead of HardNet descriptor | HardNet · Heinly benchmark, Exposure | matching score: 34.9 -> 36.3 | Improved |
| use GeoDesc instead of HardNet descriptor | HardNet · Heinly benchmark, Planar | matching score: 27.4 -> 29.1 | Improved |
| use GeoDesc instead of DSP-SIFT | DSP-SIFT · Heinly benchmark, JPEG and Blur | matching score: 35.1 vs 34.7 (JPEG), 14.3 vs 14.4 (Blur) | No clear effect |
Later papers that compared with this method
1 later accepted paper reports GeoDesc as a baseline in their results tables; 1 of them report a better number on most of the results they share with it.
- Learning Feature Descriptors using Camera Pose Supervision (ECCV 2020 · better on 1 of 1 shared result)
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
- Learning 3D Keypoint Descriptors for Non-Rigid Shape Matching (ECCV 2018 · similarity 0.87)
- 3DMatch: Learning Local Geometric Descriptors From RGB-D Reconstructions (CVPR 2017 · similarity 0.87)
- Geodesic-HOF: 3D Reconstruction Without Cutting Corners (AAAI 2021 · similarity 0.87)
- From Extrinsic to Intrinsic: Geodesic-Guided Representation Learning for 3D Geometric Data (ICML 2026 · similarity 0.86)
- LP-DIF: Learning Local Pattern-Specific Deep Implicit Function for 3D Objects and Scenes (CVPR 2023 · similarity 0.86)
- DGC-GNN: Leveraging Geometry and Color Cues for Visual Descriptor-Free 2D-3D Matching (CVPR 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.
