Research map / Detection and segmentation / CNN features
SuperJunction: Learning-Based Junction Detection for Retinal Image Registration
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.
Proposes SuperJunction, a self-supervised junction detector and descriptor for retinal image registration, using vessel-detection multi-task regularization, constrained negative sampling, and hybrid homography/polynomial matching.
Problem
Keypoint-based retinal image registration suffers from ineffective keypoint detector and descriptor training and ignores the non-linear mapping between differing views of the 3D retina.
Approach
Detect vessel junctions (bifurcations/crossovers) as keypoints with a SuperPoint-style ResNet-34 network regularized by an auxiliary vessel-segmentation decoder, train the descriptor with a constrained negative sampling loss that picks negatives from the top-K most similar keypoints, and switch between homography and second-order polynomial mapping depending on translation magnitude.
Key results
- SuperJunction reaches mAUC 0.850 on FIRE, 12.6% above SuperRetina's 0.755.
- Subset P AUC improves from 0.542 (SuperRetina) to 0.802, over 47.9% gain.
- Ablation: vessel regularization raises keypoint recall from 47.7% to 55.2%, and CNS raises it further to 63.5% at equal precision (63.9%).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| use SuperJunction instead of SuperRetina | Retinal image registration · FIRE | mAUC: 0.755 -> 0.850 | Helped |
| use SuperJunction instead of REMPE | Retinal image registration · FIRE subset S | AUC-S: 0.958 -> 0.954 | Hurt |
| add vessel-detection regularization (ablation) | SuperJunction · Retinal keypoint test images | recall: 47.7% -> 55.2% | Helped |
| add constrained negative sampling (CNS) after vessel regularization (ablation) | SuperJunction · Retinal keypoint test images | recall: 55.2% -> 63.5% | Helped |
| add vessel-detection regularization (ablation) | SuperJunction · FIRE | mAUC: 0.809 -> 0.839 | Helped |
| add CNS after vessel regularization (ablation) | SuperJunction · FIRE | mAUC: 0.839 -> 0.850 | Helped |
| add vessel-detection regularization (ablation) | SuperJunction · FIRE subset A | AUC-A: 0.780 -> 0.805 | Helped |
| add CNS after vessel regularization (ablation) | SuperJunction · FIRE subset A | AUC-A: 0.805 -> 0.794 | Hurt |
| use second-order rather than third-order polynomial fitting (ablation) | Hybrid matching · FIRE subset P | AUC-P: 0.798 -> 0.802 | Helped |
| use second-order rather than fourth-order polynomial fitting (ablation) | Hybrid matching · FIRE subset P | AUC-P: 0.801 -> 0.802 | Helped |
| use second-order rather than fifth-order polynomial fitting (ablation) | Hybrid matching · FIRE subset P | AUC-P: 0.782 -> 0.802 | Helped |
Surprising findings
- Plugging polynomial fitting into SuperRetina did not help (AUC 0.539 for subset P vs 0.542 without), attributed to mismatched keypoints.
Limitations
- May not work well when disease substantially alters the vessel pattern.
- Polynomial fitting depends on accurately matched points; errors in matching degrade it.
Experimental setup
Datasets: DRIVE STARE HRF CHASE_DB FIRE
Benchmarks: FIRE
Metrics: precision recall AUC mAUC RMSE
Baselines: SIFT PBO REMPE SuperPoint GLAMpoints R2D2 SuperGlue NCNet SuperRetina
Builds on: SuperPoint SuperGlue SuperRetina GLAMpoint
Closest papers by meaning
- R2D2: Reliable and Repeatable Detector and Descriptor (NeurIPS 2019 · similarity 0.82)
- Dense Keypoints via Multiview Supervision (NeurIPS 2021 · similarity 0.82)
- FOCUS: Benchmarking Retinal Model Generalization from Foundation Vision Encoders to Multimodal LLMs (NeurIPS 2026 · similarity 0.82)
- SPGroup3D: Superpoint Grouping Network for Indoor 3D Object Detection (AAAI 2024 · similarity 0.82)
- Improved Efficiency Based on Learned Saccade and Continuous Scene Reconstruction From Foveated Visual Sampling (ICLR 2024 · similarity 0.82)
- Incremental Object Keypoint Learning (CVPR 2025 · similarity 0.81)
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.