Research map / Label-efficient and robust learning / Recognition and long-tail
Tattoo Image Search at Scale: Joint Detection and Compact Representation Learning
Hu Han, Jie Li, Anil K. Jain, Shiguang Shan, Xilin Chen
Read the paperCheck your idea against 137,217 papers
AI-extracted. The measured effects below were read by AI from the paper's full text (no summary yet). Numbers and claims can be wrong or missing context, so check the paper before you cite or build on it.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add joint detection and compact representation learning (multi-task) on top of a Faster R-CNN detector | tattoo detector · Tatt-C | recall at 0.1 FPPI: 56.2 -> 61.7 | Improved |
| add joint detection and compact representation learning (multi-task) | tattoo detector · WebTattoo test | recall at 0.1 FPPI: 80.2 -> 87.1 | Improved |
| detect tattoos with Faster R-CNN before SSDH features (ablation) | SSDH-VGG16 · WebTattoo search | precision-recall | Improved |
| use ground-truth tattoo boxes instead of detected boxes for features | proposed CRL · WebTattoo search | precision-recall | Improved |
| extract features from final location estimates instead of region proposals | proposed vs OIM-ResNet50 · WebTattoo search | precision-recall | Improved |
| increase code length from 256 to 512 bits (ablation) | proposed · WebTattoo search | precision: minor precision drop around 60% recall | No clear effect |
| reduce code length from 256 to 128 bits (ablation) | proposed · WebTattoo search | precision: large drop of precision | Hurt |
| add dispersity loss to cross-entropy loss (ablation) | proposed · WebTattoo search | rank-1 accuracy: no rank-1 gain, improves beyond rank-1 | No clear effect |
| add polarization loss to cross-entropy and dispersity losses (ablation) | proposed · WebTattoo search | rank-1 accuracy | Improved |
| populate gallery with 300K background tattoo images | all methods · WebTattoo identification | rank-1 accuracy: about 30% degradation at rank-1 | Hurt |
| use binary codes with Hamming distance instead of real-valued cosine distance | proposed · 256-D codes | matching time (ms): 0.3 -> 0.06 | Improved |
| stitch multiple training images and buffer features to enlarge instance batch for CRL | proposed training · tattoo search | training convergence | Improved |
Closest papers by meaning
- Tasks Integrated Networks: Joint Detection and Retrieval for Image Search (TPAMI 2020 · similarity 0.82)
- Semi-Supervised Skin Detection by Network With Mutual Guidance (ICCV 2019 · similarity 0.80)
- One for All: Synthesis-Free Fingerprint Learning for Attribution of In-the-Wild Synthetic Images (AAAI 2026 · similarity 0.80)
- Optimal Transport-based Labor-free Text Prompt Modeling for Sketch Re-identification (NeurIPS 2024 · similarity 0.80)
- ARNet: Self-Supervised FG-SBIR with Unified Sample Feature Alignment and Multi-Scale Token Recycling (AAAI 2025 · similarity 0.79)
- Community Forensics: Using Thousands of Generators to Train Fake Image Detectors (CVPR 2025 · similarity 0.78)
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.
