Research map / 3D vision / Depth and stereo
RUBIK: A Structured Benchmark for Image Matching across Geometric Challenges
Read the paperCheck your idea against 108,967 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.
Introduces RUBIK, a benchmark of 16.5K nuScenes image pairs organized into 33 difficulty levels by overlap, scale ratio, and viewpoint angle, and evaluates 14 image matching methods on it.
Problem
Existing image matching and camera pose benchmarks offer limited insight into how methods fail under distinct geometric challenges such as low overlap, large scale change, and large viewpoint change.
Approach
Defines three geometric difficulty criteria (overlap, median scale ratio, median viewpoint angle) computed from dense co-visibility maps built on lifted 6-DoF nuScenes poses and monocular depth/normal maps, then bins image pairs into a grid of 33 valid difficulty levels with 500 pairs each.
Key results
- Best method DUSt3R reaches 54.8% overall success rate, with MASt3R at 53.6% and RoMa at 47.3%.
- Detector-free methods run in roughly 150-600 ms per pair versus about 40-70 ms for the fastest detector-based methods (ALIKED+LightGlue 45 ms, SP+LightGlue 43 ms, DISK+LightGlue 69 ms).
- The benchmark contains 16.5K pairs across 33 difficulty levels, with 500 pairs per level.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use detector-free matching instead of detector-based matching | Image matching methods on RUBIK · RUBIK | success rate: >47% vs. 14.2–36.8% | Helped |
| Use detector-free matching instead of detector-based matching | Image matching methods on RUBIK · RUBIK | runtime: 150-600ms vs. 40-70ms | Hurt |
| Use XFeat+LighterGlue instead of XFeat | XFeat on RUBIK · RUBIK | success rate: 14.2 -> 30.1 | Helped |
| Use XFeat+LighterGlue instead of XFeat | XFeat on RUBIK · RUBIK | runtime: 54 -> 43 | Helped |
Surprising findings
- Even the best method fails on more than 45% of pairs, indicating substantial remaining difficulty.
- XFeat with LighterGlue performs much better than plain XFeat, showing the matching strategy matters even with the same detector.
Limitations
- Co-visibility pipeline is not robust to instance changes, such as a car replaced by another car, which can be wrongly labeled co-visible.
- Restricted to good-weather nuScenes scenes, so adverse conditions are not covered.
Experimental setup
Datasets: nuScenes
Benchmarks: RUBIK
Metrics: success rate (rotation error < 5°, translation error < 2m) median runtime per image pair (ms) average rank
Baselines: SIFT+LightGlue SuperPoint+LightGlue DISK+LightGlue ALIKED+LightGlue DeDoDe v2 XFeat XFeat* XFeat+LighterGlue LoFTR ELoFTR ASpanFormer RoMa
Builds on: COLMAP MAGSAC++ UniDepth Depth Anything V2
Closest papers by meaning
- UNOPose: Unseen Object Pose Estimation with an Unposed RGB-D Reference Image (CVPR 2025 · similarity 0.87)
- RefPose: Leveraging Reference Geometric Correspondences for Accurate 6D Pose Estimation of Unseen Objects (CVPR 2025 · similarity 0.86)
- SingRef6D: Monocular Novel Object Pose Estimation with a Single RGB Reference (NeurIPS 2025 · similarity 0.86)
- From Correspondences to Pose: Non-minimal Certifiably Optimal Relative Pose without Disambiguation (CVPR 2024 · similarity 0.85)
- Depth2Pose: A Pose-Based Benchmark for Monocular Depth Estimation without Ground-Truth Depth (NeurIPS 2026 · similarity 0.85)
- HomoMatcher: Achieving Dense Feature Matching with Semi-Dense Efficiency by Homography Estimation (AAAI 2025 · 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.