mlsanityData to 14 Oct 2026

RoMa: Robust Dense Feature Matching

arXiv 2305.15404 ↗First posted 24 May 2023cs.CVcited by 302

Not on the map. Why?

Abstract (arXiv)

Feature matching is an important computer vision task that involves estimating correspondences between two images of a 3D scene, and dense methods estimate all such correspondences. The aim is to learn a robust model, i.e., a model able to match under challenging real-world changes. In this work, we propose such a model, leveraging frozen pretrained features from the foundation model DINOv2. Although these features are significantly more robust than local features trained from scratch, they are inherently coarse. We therefore combine them with specialized ConvNet fine features, creating a precisely localizable feature pyramid. To further improve robustness, we propose a tailored transformer match decoder that predicts anchor probabilities, which enables it to express multimodality. Finally, we propose an improved loss formulation through regression-by-classification with subsequent robust regression. We conduct a comprehensive set of experiments that show that our method, RoMa, achieves significant gains, setting a new state-of-the-art. In particular, we achieve a 36% improvement on the extremely challenging WxBS benchmark. Code is provided at https://github.com/Parskatt/RoMa

Key facts. Cited by 302 papers in the map (data through 14 Oct 2026). Builds on 46 of its 69 references that are in the map. First posted on arXiv 24 May 2023.

Builds on

46 of 69 references.

References · 46

  1. 2015Deep Residual Learning for Image Recognitioncited by 52,096
  2. 2018BERT: Pre-training of Deep Bidirectional Transformers for Language Understandingcited by 40,399
  3. 2023DINOv2: Learning Robust Visual Features without Supervisioncited by 7,974
  4. 2021Masked Autoencoders Are Scalable Vision Learnerscited by 7,403
  5. 2021Emerging Properties in Self-Supervised Vision Transformerscited by 6,968
  6. 2021On the Opportunities and Risks of Foundation Modelscited by 4,033
  7. 2017ScanNet: Richly-annotated 3D Reconstructions of Indoor Scenescited by 3,117
  8. 2019SuperGlue: Learning Feature Matching with Graph Neural Networkscited by 1,443
  9. 2017SuperPoint: Self-Supervised Interest Point Detection and Descriptioncited by 1,320
  10. 2021LoFTR: Detector-Free Local Feature Matching with Transformerscited by 984
  11. 2021iBOT: Image BERT Pre-Training with Online Tokenizercited by 903
  12. 2018MegaDepth: Learning Single-View Depth Prediction from Internet Photoscited by 737
34 more
  1. 2021Masked Feature Prediction for Self-Supervised Visual Pre-Trainingcited by 576
  2. 2018From Coarse to Fine: Robust Hierarchical Localization at Large Scalecited by 559
  3. 2020DISK: Learning local features with policy gradientcited by 312
  4. 2020Prior Guided Feature Enrichment Network for Few-Shot Segmentationcited by 271
  5. 2017HPatches: A benchmark and evaluation of handcrafted and learned local descriptorscited by 264
  6. 2017A General and Adaptive Robust Loss Functioncited by 236
  7. 2021Back to the Feature: Learning Robust Camera Localization from Pixels to Posecited by 199
  8. 2022DKM: Dense Kernelized Feature Matching for Geometry Estimationcited by 177
  9. 2018Neighbourhood Consensus Networkscited by 174
  10. 2022ASpanFormer: Detector-Free Image Matching with Adaptive Span Transformercited by 171
  11. 2019MAGSAC++, a fast, reliable and accurate robust estimatorcited by 161
  12. 2019GLU-Net: Global-Local Universal Network for Dense Flow and Correspondencescited by 139
  13. 2022QuadTree Attention for Vision Transformerscited by 129
  14. 2022MatchFormer: Interleaving Attention in Transformers for Feature Matchingcited by 126
  15. 2022Revealing the Dark Secrets of Masked Image Modelingcited by 117
  16. 2021Learning Accurate Dense Correspondences and When to Trust Themcited by 113
  17. 2019Hierarchical Scene Coordinate Classification and Regression for Visual Localizationcited by 93
  18. 2020Dual-Resolution Correspondence Networkscited by 90
  19. 2021PDC-Net+: Enhanced Probabilistic Dense Correspondence Networkcited by 80
  20. 2018DGC-Net: Dense Geometric Correspondence Networkcited by 67
  21. 2021CamLiFlow: Bidirectional Camera-LiDAR Fusion for Joint Optical Flow and Scene Flow Estimationcited by 56
  22. 2023SiLK -- Simple Learned Keypointscited by 54
  23. 2020GOCor: Bringing Globally Optimized Correspondence Volumes into Your Neural Networkcited by 53
  24. 2020Wasserstein Distances for Stereo Disparity Estimationcited by 38
  25. 2023PATS: Patch Area Transportation with Subdivision for Local Feature Matchingcited by 33
  26. 2023PMatch: Paired Masked Image Modeling for Dense Geometric Matchingcited by 29
  27. 2022A case for using rotation invariant features in state of the art feature matcherscited by 27
  28. 2015WxBS: Wide Baseline Stereo Generalizationscited by 23
  29. 2023Adaptive Spot-Guided Transformer for Consistent Local Feature Matchingcited by 22
  30. 2022Proper Reuse of Image Classification Features Improves Object Detectioncited by 20
  31. 2021Neural Reprojection Error: Merging Feature Learning and Camera Pose Estimationcited by 17
  32. 2023Improving Transformer-based Image Matching by Cascaded Capturing Spatially Informative Keypointscited by 15
  33. 2019Large Scale Joint Semantic Re-Localisation and Scene Understanding via Globally Unique Instance Coordinate Regressioncited by 12
  34. 2022Could Giant Pretrained Image Models Extract Universal Representations?cited by 6

Built on by

Cited by year (2026 to 14 Oct)

Citing papers per year: 2023: 4; 2024: 39; 2025: 130; 2026: 129.

302 papers.

Sort:

By area:

  1. 2023DUSt3R: Geometric 3D Vision Made Easycited by 1,567
  2. 2025VGGT: Visual Geometry Grounded Transformercited by 1,539
  3. 2024Grounding Image Matching in 3D with MASt3Rcited by 983
  4. 2024No Pose, No Problem: Surprisingly Simple 3D Gaussian Splats from Sparse Unposed Imagescited by 248
  5. 2024MASt3R-SfM: a Fully-Integrated Solution for Unconstrained Structure-from-Motioncited by 162
  6. 2025Aether: Geometric-Aware Unified World Modelingcited by 97
  7. 2024Reloc3r: Large-Scale Training of Relative Camera Pose Regression for Generalizable, Fast, and Accurate Visual Localizationcited by 82
  8. 2024DINOv2 Meets Text: A Unified Framework for Image- and Pixel-Level Vision-Language Alignmentcited by 79
  9. 2025Scenethesis: A Language and Vision Agentic Framework for 3D Scene Generationcited by 67
  10. 2024PF3plat: Pose-Free Feed-Forward 3D Gaussian Splattingcited by 62
  11. 2025FreeTimeGS: Free Gaussian Primitives at Anytime and Anywhere for Dynamic Scene Reconstructioncited by 54
  12. 2025RoMa v2: Harder Better Faster Denser Feature Matchingcited by 51
38 more
  1. 2025Light3R-SfM: Towards Feed-forward Structure-from-Motioncited by 47
  2. 2024SelfSplat: Pose-Free and 3D Prior-Free Generalizable 3D Gaussian Splattingcited by 42
  3. 2024FreeSplatter: Pose-free Gaussian Splatting for Sparse-view 3D Reconstructioncited by 41
  4. 2025No Pose at All: Self-Supervised Pose-Free 3D Gaussian Splatting from Sparse Viewscited by 40
  5. 2024DeDoDe v2: Analyzing and Improving the DeDoDe Keypoint Detectorcited by 39
  6. 2024CRAB: Cross-environment Agent Benchmark for Multimodal Language Model Agentscited by 39
  7. 2025AerialMegaDepth: Learning Aerial-Ground Reconstruction and View Synthesiscited by 39
  8. 2025YoNoSplat: You Only Need One Model for Feedforward 3D Gaussian Splattingcited by 37
  9. 2025UFM: A Simple Path towards Unified Dense Correspondence with Flowcited by 36
  10. 2024MVInpainter: Learning Multi-View Consistent Inpainting to Bridge 2D and 3D Editingcited by 30
  11. 2024Cross-View Completion Models are Zero-shot Correspondence Estimatorscited by 30
  12. 2025MP-SfM: Monocular Surface Priors for Robust Structure-from-Motioncited by 26
  13. 2025Emergent Outlier View Rejection in Visual Geometry Grounded Transformerscited by 22
  14. 2025C3G: Learning Compact 3D Representations with 2K Gaussianscited by 21
  15. 2025iLRM: An Iterative Large 3D Reconstruction Modelcited by 19
  16. 2024FewViewGS: Gaussian Splatting with Few View Matching and Multi-stage Trainingcited by 18
  17. 2025Relative Pose Estimation through Affine Corrections of Monocular Depth Priorscited by 18
  18. 2025Cluster and Predict Latent Patches for Improved Masked Image Modelingcited by 17
  19. 2024How Far are AI-generated Videos from Simulating the 3D Visual World: A Learned 3D Evaluation Approachcited by 16
  20. 2023Segment Anything Model is a Good Teacher for Local Feature Learningcited by 15
  21. 2025Dense-SfM: Structure from Motion with Dense Consistent Matchingcited by 15
  22. 2025HAMSt3R: Human-Aware Multi-view Stereo 3D Reconstructioncited by 15
  23. 2025JamMa: Ultra-lightweight Local Feature Matching with Joint Mambacited by 14
  24. 2025Co-op: Correspondence-based Novel Object Pose Estimationcited by 14
  25. 2025From Sparse to Dense: Camera Relocalization with Scene-Specific Detector from Feature Gaussian Splattingcited by 14
  26. 2025DiffusionSfM: Predicting Structure and Motion via Ray Origin and Endpoint Diffusioncited by 14
  27. 2025ThermalGen: Style-Disentangled Flow-Based Generative Models for RGB-to-Thermal Image Translationcited by 14
  28. 2024LoD-Loc: Aerial Visual Localization using LoD 3D Map with Neural Wireframe Alignmentcited by 13
  29. 2025DaD: Distilled Reinforcement Learning for Diverse Keypoint Detectioncited by 13
  30. 2025DreamFuse: Adaptive Image Fusion with Diffusion Transformercited by 13
  31. 2025RDD: Robust Feature Detector and Descriptor using Deformable Transformercited by 13
  32. 2025MATCHA:Towards Matching Anythingcited by 12
  33. 2025CoMatch: Dynamic Covisibility-Aware Transformer for Bilateral Subpixel-Level Semi-Dense Image Matchingcited by 12
  34. 2025OrthoLoC: UAV 6-DoF Localization and Calibration Using Orthographic Geodatacited by 12
  35. 2025Rig3R: Rig-Aware Conditioning for Learned 3D Reconstructioncited by 11
  36. 2026LoMa: Local Feature Matching Revisitedcited by 11
  37. 2023TopicFM+: Boosting Accuracy and Efficiency of Topic-Assisted Feature Matchingcited by 10
  38. 2025RePoseD: Efficient Relative Pose Estimation With Known Depth Informationcited by 10

50 most cited of 302 All 302 (JSON)