Research map / Detection and segmentation / Open-set and few-shot detection
Spotting the Unseen: Reciprocal Consensus Network Guided by Visual Archetypes
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 CoNet, a one-shot detection/spotting network that matches query image regions to visual archetypes (support images) via auto-correlation and mutual-correlation representations with reciprocal consistency and an adaptive margin contrastive loss.
Problem
One-shot/few-shot object and text spotting of unseen categories from a few visual archetypes, especially in low-data regimes and for confusable classes such as historical manuscript characters and natural scene text.
Approach
Combines intra-image auto-correlation (local self-similarity plus global context) with inter-image mutual correlation filtered by a reciprocal (bidirectional) consistency constraint, followed by geometric alignment via affine grid sampling, and trains with a contrastive loss using adaptive margins per sample.
Key results
- On ICDAR-13 unseen classes, CoNet surpasses the runner-up mAP by 30.7%.
- On Paste-f, CoNet reaches 58.27 mAP / 56.88 F1 versus OS2D's 52.56 mAP / 47.76 F1.
- Ablation on Paste-f: removing ACR drops F1 by 7.07 points and removing the reciprocal consistency constraint drops it by 7.21 points.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Use CoNet instead of the strongest baseline OS2D | object detection · Paste-f | F1: 47.76 -> 56.88 | Helped |
| Use CoNet instead of CoAE | logo detection · FlickrLogos-27 seen | F1: 28.03 -> 24.14 | Hurt |
| Add ACR (auto-correlation representation) (ablation) | CoNet · KAIST-E unseen | F1: 88.39 -> 92.42 | Helped |
| Add ACR (auto-correlation representation) (ablation) | CoNet · KAIST-K unseen | F1: 86.21 -> 91.75 | Helped |
| Add reciprocal consistency constraint (ablation) | CoNet · Paste-f unseen | F1: 49.67 -> 56.88 | Helped |
| Add reciprocal consistency constraint (ablation) | CoNet · KAIST-E unseen | F1: 87.55 -> 92.42 | Helped |
| Replace fixed margins with adaptive margin strategy (ablation) | CoNet · Dairy (unseen) | F1: 78.99 -> 81.66 | Helped |
| Replace fixed margins with adaptive margin strategy (ablation) | CoNet · KAIST-E unseen | F1: 69.46 -> 92.42 | Helped |
| Increase training images or categories (ablation) | CoNet on TKH unseen categories · TKH unseen categories | F1 | Helped |
Surprising findings
- CoAE performs better than CoNet on some seen categories of FlickrLogos-27 and on seen classes of several datasets, and unsupervised DSW surpasses supervised CoAE on TKH unseen classes.
Limitations
- CoNet does not achieve best results on unseen classes of GroZi-3.2k and does not lead on some seen-category metrics.
- Logo detection results on FlickrLogos-27 remain low, attributed by the authors to low-quality visual archetypes.
Experimental setup
Datasets: GroZi-3.2k Dairy Paste FlickrLogos-32 Dongba Hieroglyphics (DBH) VML-HD Tripitaka Koreana in Han (TKH) ICDAR-13 KAIST
Benchmarks: GroZi-3.2k Dairy Paste-v Paste-f FlickrLogos-27 DBH VML-HD TKH ICDAR-13 KAIST-E KAIST-K
Metrics: mAP recall F1
Baselines: DSW CoAE OS2D
Builds on: OS2D CoAE Dynamic Context Correspondence Network Ranked List Loss CBAM
Closest papers by meaning
- Correspondence Networks With Adaptive Neighbourhood Consensus (CVPR 2020 · similarity 0.86)
- MR-NET: Exploiting Mutual Relation for Visual Relationship Detection (AAAI 2019 · similarity 0.86)
- CoADNet: Collaborative Aggregation-and-Distribution Networks for Co-Salient Object Detection (NeurIPS 2020 · similarity 0.85)
- Visual Consensus Prompting for Co-Salient Object Detection (CVPR 2025 · similarity 0.85)
- CLIP in Mirror: Disentangling text from visual images through reflection (NeurIPS 2024 · similarity 0.84)
- BCLNet: Bilateral Consensus Learning for Two-View Correspondence Pruning (AAAI 2024 · similarity 0.84)
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.