Research map / Label-efficient and robust learning / Domain adaptation and generalization
Mixture of Submodules for Domain Adaptive Person Search
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 MoS, a mixture-of-submodules framework built on DAPS that routes task- and domain-specific submodules for domain adaptive person search, plus counterpart domain sample generation via denormalization with contrastive domain alignment.
Problem
Domain adaptive person search jointly requires person detection and re-identification, whose contradictory objectives cause gradient conflict when sharing parameters in a unified framework across domains.
Approach
Decompose the norm-aware embedding head into submodules by dropping convolutional blocks, and use a conditional top-K router driven by task and domain embeddings so detection and ReID (and source/target) use different submodule mixtures; synthesize counterpart-domain ReID samples by swapping BN statistics (denormalization) and train with a contrastive domain alignment loss.
Key results
- On CUHK-SYSU to PRW, MoS reaches 37.1% mAP and 81.9% top-1, improving over DAPS by 2.4 and 1.3 points.
- On PRW to CUHK-SYSU, MoS reaches 80.1% mAP and 81.5% top-1, outperforming DAPS by 2.5 and 1.9 points.
- Ablation: task-conditioned routing raises PRW mAP from 34.7 to 35.3; adding domain conditioning gives 36.0; counterpart generation gives 37.1.
Measured effects reported in the paper
| What was changed | Where | Measured | Effect |
|---|---|---|---|
| replace DAPS with MoS | DAPS · CUHK-SYSU → PRW | mAP: 34.7 -> 37.1 | Helped |
| replace DAPS with MoS | DAPS · CUHK-SYSU → PRW | top-1 accuracy: 80.6 -> 81.9 | Helped |
| replace DAPS with MoS | DAPS · PRW → CUHK-SYSU | mAP: 77.6 -> 80.1 | Helped |
| replace DAPS with MoS | DAPS · PRW → CUHK-SYSU | top-1 accuracy: 79.6 -> 81.5 | Helped |
| add task-conditional mixture of submodules (ablation) | DAPS · CUHK-SYSU → PRW | mAP: 34.7 -> 35.3 | Helped |
| add domain-conditional routing to task-conditional mixture of submodules (ablation) | MoS (T) · CUHK-SYSU → PRW | mAP: 35.3 -> 36.0 | Helped |
| add counterpart domain sample generation (ablation) | MoS (T + D) · CUHK-SYSU → PRW | mAP: 36.0 -> 37.1 | Helped |
| set Top-K from 1 to 2 (ablation) | MoS · CUHK-SYSU → PRW | mAP: 35.9 -> 37.1 | Helped |
| increase Top-K from 2 to 3, 4, or 5 (ablation) | MoS · CUHK-SYSU → PRW | mAP: 37.1; 36.8; 36.1; 35.4 | Hurt |
| increase negative samples from 16 to 128 (ablation) | MoS contrastive domain alignment · CUHK-SYSU → PRW | mAP: 36.2; 36.5; 36.8; 37.1 | Helped |
| increase negative samples from 128 to 256 (ablation) | MoS contrastive domain alignment · CUHK-SYSU → PRW | mAP: 37.1 -> 36.9 | Hurt |
Surprising findings
- Increasing top-K beyond 2 degraded performance, which the authors attribute to more submodules being shared across domains and tasks.
Limitations
- Gains over DAPS are moderate (2.4 mAP on PRW, 2.5 mAP on CUHK-SYSU).
- Results reported only on two domain pairs with ResNet50 backbone.
Experimental setup
Datasets: CUHK-SYSU PRW
Benchmarks: CUHK-SYSU PRW
Metrics: mAP top-1 accuracy average precision recall rate
Baselines: DAPS OIM MGTS IGPN TCTS NAE AlignPS SeqNet
Builds on: DAPS OIM loss domain-specific batch normalization top-K gating
Closest papers by meaning
- Localization-Anchored Instance Discrimination for Domain Adaptive Person Search (AAAI 2026 · similarity 0.89)
- Doubly Contrastive Learning for Source-Free Domain Adaptive Person Search (AAAI 2025 · similarity 0.88)
- Invariance Matters: Exemplar Memory for Domain Adaptive Person Re-Identification (CVPR 2019 · similarity 0.88)
- Unsupervised Domain Adaptive Person Search via Dual Self-Calibration (AAAI 2025 · similarity 0.87)
- Divide and Conquer: Hybrid Pre-training for Person Search (AAAI 2024 · similarity 0.86)
- Domain Generalizable Person Search Using Unreal Dataset (AAAI 2024 · similarity 0.85)
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.