Research map / Label-efficient and robust learning / Recognition and long-tail
Differentiable Auxiliary Learning for Sketch Re-Identification
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 DALNet, which generates a sketch-like auxiliary modality from photos with a learnable dynamic auxiliary generator and aligns photo, sketch and auxiliary features with modality interactive attention and circle-loss-based collaborative learning for Sketch Re-ID.
Problem
Sketch re-identification must match pedestrian photos to sketches despite large cross-modality gaps, background clutter, light variation, varied sketch styles, and limited samples per modality that leave latent feature distributions poorly constrained.
Approach
Introduces a dynamically trained auxiliary modality (sketch-like version of photos produced by a lightweight U2-Net plus learnable edge-filter generator, refined by a style contrastive loss) and uses it as a bridge via modality interactive attention and a cross-modality plus intra-modality circle loss.
Key results
- On PKU-Sketch, DALNet reaches Rank-1 90.0% and mAP 86.2%, versus CCSC at 86.0% Rank-1 and 83.7% mAP.
- On ChairV2, Rank-1 82.3% and Rank-10 98.3%; on ShoeV2, Rank-1 48.5% and Rank-10 86.1%.
- On CUHK student, Rank-1 96.50%, 2.5 points above CDAC; on CUFSF, Rank-1 97.84%.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| Add DALNet’s auxiliary modality and cross-modality circle loss (ablation) | ResNet-50 identity-loss baseline · PKU-Sketch | Rank-1: 67.8 -> 76.5 | Helped |
| Replace original circle loss with cross-modality circle loss (ablation) | Auxiliary-modality model · PKU-Sketch | Rank-1: 71.6 -> 76.5 | Helped |
| Add modality interactive attention (ablation) | Auxiliary-modality model with cross-modality circle loss · PKU-Sketch | Rank-1: 76.5 -> 86.4 | Helped |
| Add intra-modality circle loss (ablation) | DALNet without intra-modality circle loss · PKU-Sketch | Rank-1: 86.4 -> 90.0 | Helped |
| Add intra-modality circle loss (ablation) | DALNet without intra-modality circle loss · ShoeV2 | Rank-1: 45.8 -> 48.5 | Helped |
| Jointly optimize the auxiliary generator with style refinement learning (ablation) | DALNet generator · PKU-Sketch | Rank-1: 86.5 -> 90.0 | Helped |
| Jointly optimize the auxiliary generator with style refinement learning (ablation) | DALNet generator · ShoeV2 | Rank-1: 45.2 -> 48.5 | Helped |
Limitations
- Gains over the strongest baseline on PKU-Sketch are modest (about 4 points Rank-1), and ShoeV2 comparisons rely on a single run.
- Relies on a pre-trained U2-Net for foreground separation, so generator quality depends on that external model.
Experimental setup
Datasets: PKU-Sketch QMUL-ShoeV2 QMUL-ChairV2 CUHK student CUFSF
Benchmarks: PKU-Sketch QMUL-ShoeV2 QMUL-ChairV2 CUHK student CUFSF
Metrics: CMC Rank-1 Rank-5 Rank-10 mAP
Baselines: AFLNet LMDI CDAC SketchTrans CCSC Sketch-PVT Stylemeup KD IA CycleGAN MMTN
Builds on: circle loss SketchTrans U2-Net ResNet-50
Closest papers by meaning
- Optimal Transport-based Labor-free Text Prompt Modeling for Sketch Re-identification (NeurIPS 2024 · similarity 0.88)
- MDReID: Modality-Decoupled Learning for Any-to-Any Multi-Modal Object Re-Identification (NeurIPS 2025 · similarity 0.86)
- FlexiReID: Adaptive Mixture of Expert for Multi-Modal Person Re-Identification (ICML 2025 · similarity 0.86)
- Implicit Discriminative Knowledge Learning for Visible-Infrared Person Re-Identification (CVPR 2024 · similarity 0.85)
- Multi-Modal Object Re-identification via Sparse Mixture-of-Experts (ICML 2025 · similarity 0.85)
- ReID5o: Achieving Omni Multi-modal Person Re-identification in a Single Model (NeurIPS 2025 · 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.