검색 상세

Background Aware Contrastive Learning in Domain Adaptive Person Search

  • 전창렬 (서강대학교 일반대학원)

초록

The person search aims to jointly address Pedestrian Detection and Person Re- identification (re-ID). However, in domain shift scenarios, models for this task fre- quently experience significant performance degradation. However, current research on domain adaptive person search is not optimized for these challenges. It relies on a combination of domain adaptation techniques from pedestrian detection and re- ID. In this paper, we address this issue by proposing a solution for domain adaptive person search tasks that merges the Teacher-Student structure with contrastive learn- ing, a method we call Background Aware Contrastive Learning (BACL). Initially, the Teacher-Student structure facilitates the use of various feature representations for training while incorporating precise proposal boxes. Our contrastive learning strat- egy effectively integrates background clutter, person instance features, and source domain memory, fully leveraging the dataset. Moreover, Our method optimizes per- son search by simultaneously classifying person instances and background clutter, and enhancing re-ID training. We accomplish this goal by utilizing diverse positive pairs, which allows for the learning of more varied and robust characteristics for a single person instance. Our experimental results on the PRW and CUHK-SYSU datasets demonstrate our approach’s compatibility with existing methods.

more

초록

인물 검색은 보행자 감지와 인물 재식별(re-ID)을 함께 해결하려는 목표를 가지고있다. 하지만,도메인변화상황에서이작업을위한모델들은종종성능저하를 겪는다. 현재의 도메인 적응형 인물 검색 연구는 이러한 도전 과제에 특화되어 최적화 되지 않고,보행자감지와객체감지에서사용되는도메인적응기술의조합 에 의존한다. 이 논문에서는 ’배경 인식 대조 검색(Background Aware Contrastive Learning, BACL)’이라고부르는방법으로,교사-학생구조와대조학습을결합한 도메인적응형인물검색작업을위한해결책을제안한다.우선교사-학생 구조는 다양한 특징 표현을 훈련에 사용하면서 정확한 제안 상자를 통합하는 데 도움을 준다. 다음으로, 대조 학습 전략을 사용함으로써, 배경 잡음, 인물 인스턴스 특징, 그리고원본도메인메모리를효과적으로통합하여데이터셋을충분히활용할수 있게 한다. 또한,우리의학습방법은배경잡음과개인인스턴스를분류하는학습을 진행함으로써,개별인스턴스에대한다양한긍정적쌍의사용을통해재식별 성능을 향상 시키는 인물 검색 작업을 위한 최적화된 해결책을 제공한다. PRW와 CUHK-SYSU 데이터 셋에 대한 실험 결과는 우리의 접근 방식이 기존 방법 대비 높은성능을기록하였다.

more

목차

I . Introduction 1
II . Related Works 6
2.1 Person Search 6
2.2 Unsupervised domain adaptation (UDA) for person Re-identification 7
2.3 Contrastive Learning 7
III . Method 9
3.1 Domain Adaptive Person Search 9
3.1.1 Network 9
3.1.2 Hybrid Memory 10
3.1.3 Loss 10
3.2 Background Aware Contrastive Learning . 13
3.2.1 Teacher-Student Framework 13
3.2.2 Contrastive Learning 14
3.2.3 Positive pair selection 16
3.2.4 Negative pair selection 16
IV . Experiments 18
4.1 Experimental setup 18
4.2 Datasets and Evaluation 18
4.2.1 Dataset 18
4.2.2 Evaluation Protocols 19
4.3 Implementation Details 19
4.4 Comparison with State-of-the-Art Methods 21
4.5 Ablation Studies 22
4.5.1 Comparisons With Augmentations 22
4.5.2 Effectiveness on Positive and Negative pair selection 23
4.5.3 Effectiveness on different image resolutions 24
4.6 Qualitative results 26
4.6.1 Visualization Results on the PRW Dataset 28
4.6.2 Visualization Results on the CUHK-SYSU Dataset 28
4.7 Further Experiment 29
4.7.1 Comparison with Prototype-based Contrastive Learning 29
4.7.2 Experimental result 30
V . Conclusion 32
Bibliography 33

more
반출 Meta View 목록