Research map / Label-efficient and robust learning / Domain adaptation and generalization
Graph-Based Cross-Domain Knowledge Distillation for Cross-Dataset Text-to-Image Person Retrieval
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 GCKD, an unsupervised domain adaptation method for cross-dataset text-to-image person retrieval that combines a graph-based multi-domain propagation module with contrastive momentum knowledge distillation on top of ALBEF.
Problem
Text-to-image person retrieval models trained with supervision degrade when transferred to a target dataset that has only unpaired, unlabeled images and texts, due to domain shift and the visual-textual modality gap.
Approach
Builds a dynamic cross-domain KNN graph over source/target image and text embeddings (with memory banks) and propagates features with a 2-layer GNN, while an EMA teacher generates pseudo image-text similarity targets for a cross-domain contrastive loss and a cross-domain matching loss.
Key results
- On ICFG-PEDES to RSTPReid, GCKD reaches 59.95 Rank-1 and 49.68 mAP, versus 55.00 Rank-1 and 46.18 mAP for the best source-only baseline (RaSa).
- On ICFG-PEDES to CUHK-PEDES, GCKD reaches 52.70 Rank-1 and 45.97 mAP, versus 48.65 Rank-1 for RaSa.
- On RSTPReid to ICFG-PEDES, GCKD reaches 46.40 Rank-1 versus 44.01 for APTM (source-only).
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| add CMKD to the backbone-only baseline (ablation) | GCKD · ICFG-PEDES → RSTPReid | Rank-1: 55.00 -> 58.20 | Helped |
| add CMKD to the backbone-only baseline (ablation) | GCKD · ICFG-PEDES → CUHK-PEDES | Rank-1: 48.65 -> 52.04 | Helped |
| add GMP to CMKD (ablation) | GCKD · ICFG-PEDES → RSTPReid | Rank-1: 58.20 -> 59.95 | Helped |
| add GMP to CMKD (ablation) | GCKD · ICFG-PEDES → CUHK-PEDES | Rank-1: 52.04 -> 52.70 | Helped |
| use GCKD instead of RaSa | text-to-image person retrieval · ICFG-PEDES → RSTPReid | Rank-1: 55.00 -> 59.95 | Helped |
| use GCKD instead of RaSa | text-to-image person retrieval · ICFG-PEDES → CUHK-PEDES | mAP: 42.03 -> 45.97 | Helped |
| use GCKD instead of APTM | text-to-image person retrieval · RSTPReid → ICFG-PEDES | Rank-1: 44.01 -> 46.40 | Helped |
| use GCKD instead of APTM | text-to-image person retrieval · RSTPReid → CUHK-PEDES | mAP: 38.47 -> 43.71 | Helped |
| use GCKD instead of RaSa | text-to-image person retrieval · CUHK-PEDES S → C03 | Rank-1: 70.63 -> 71.25 | Helped |
| use GCKD instead of RaSa | text-to-image person retrieval · CUHK-PEDES S → M | Rank-1: 82.57 -> 83.54 | Helped |
| use GCKD instead of RaSa | text-to-image person retrieval · CUHK-PEDES S → V | Rank-1: 85.20 -> 86.73 | Helped |
| use GCKD instead of RaSa | text-to-image person retrieval · CUHK-PEDES S → C01 | Rank-1: 77.34 -> 78.91 | Helped |
Surprising findings
- Existing single-domain text-to-image person retrieval methods degrade substantially in cross-dataset settings, and existing UDA baselines (POUF, ReCLIP) often underperform source-only single-domain methods.
Limitations
- Relies on an ALBEF backbone pretrained on the source dataset and on pseudo-labels from the teacher model whose quality is not analyzed in depth.
- Some experimental details (loss weights, threshold delta, memory size C) are under-specified; no code release is mentioned.
Experimental setup
Datasets: ICFG-PEDES RSTPReid CUHK-PEDES
Benchmarks: ICFG-PEDES RSTPReid CUHK-PEDES
Metrics: Rank-1 Rank-5 Rank-10 mAP
Baselines: IRRA APTM RaSa CFine IVT POUF ReCLIP MAN ECN ADDA SPGAN SSAN
Builds on: ALBEF momentum distillation ReCLIP POUF
Closest papers by meaning
- Causality-Inspired Invariant Representation Learning for Text-Based Person Retrieval (AAAI 2024 · similarity 0.84)
- DM-Adapter: Domain-Aware Mixture-of-Adapters for Text-Based Person Retrieval (AAAI 2025 · similarity 0.84)
- Lightweight Contrastive Distilled Hashing for Online Cross-modal Retrieval (AAAI 2025 · similarity 0.84)
- Multi-Modal Disordered Representation Learning Network for Description-Based Person Search (AAAI 2024 · similarity 0.84)
- Automatic Synthetic Data and Fine-grained Adaptive Feature Alignment for Composed Person Retrieval (NeurIPS 2025 · similarity 0.83)
- Domain Generalizable Person Search Using Unreal Dataset (AAAI 2024 · similarity 0.83)
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.