讲DUSt3R, MASt3R ？更直接的约束， Speedy MASt3R, 25?

讲LoFTR和Efficient LoFTR ？ 提速

讲RIPE？ 强化学习+RANSAC

讲MV-RoMa？pairs =》sequences

What

see standard pipeline: https://zju3dv.github.io/loftr/

see a demo: https://zju3dv.github.io/efficientloftr/

Classification

— Deep Learning Reforms Image Matching: A Survey and Outlook, 25

detector based vs detector free, see above figure.

According to https://github.com/ericzzj1989/Awesome-Image-Matching,

1. Detector Learning

   • Learning to Make Keypoints Sub-Pixel Accurate, eccv24
   • From Pairs to Sequences: Track-Aware Policy Gradients for Keypoint Detection, cvpr26
     1. optimize the Track-quality (Traq) of keypoints directly on image sequences
     2. Reinforcement Learning
     3. no code,

2. Descriptor Learning

   • PromptMID: Modal Invariant Descriptors Based on Diffusion and Vision Foundation Models for Optical-SAR Image Matching, ISPRS Journal of Photogrammetry and Remote Sensing, 25

3. Detector & Descriptor Learning

   • RIPE: Reinforcement Learning on Unlabeled Image Pairs for Robust Keypoint Extraction, iccv25
     • RL is OK for the mission
     • from unlabeled pairs only
     • no depth, no pose, no artificial augmentation required.
     • RIPE, iccv25

4. Feature Matching

   1. SuperGlue: Learning Feature Matching with Graph Neural Networks, cvpr20
   2. LightGlue: Local Feature Matching at Light Speed, iccv23
   3. LoFTR: Detector-Free Local Feature Matching with Transformers, cvpr21
      1. LoFTR, cvpr21
   4. Efficient LoFTR: Semi-Dense Local Feature Matching with Sparse-Like Speed, cvpr24
      1. Efficient LoFTR, cvpr24

   • RoMa: Robust Dense Feature Matching, cvpr24

   • MV-RoMa: From Pairwise Matching into Multi-View Track Reconstruction, cvpr26

     

   1. MambaGlue: Fast and Robust Local Feature Matching With Mamba, ICRA25
   2. Grounding Image Matching in 3D with MASt3R, eccv 2024

sparse, semi-dense vs dense