AdvMix: Adversarial Mixing Strategy for Unsupervised Domain Adaptive Object Detection
Recent object detection networks suffer from performance degradation when training data and test data are distinct in image styles and content distributions. In this paper, we propose a domain adaptive method, Adversarial Mixing (AdvMix), where the label-rich source domain and unlabeled target domai...
| Main Authors: | , , , , |
|---|---|
| Format: | Article |
| Language: | English |
| Published: |
MDPI AG
2024-02-01
|
| Series: | Electronics |
| Subjects: | |
| Online Access: | https://www.mdpi.com/2079-9292/13/4/685 |
| _version_ | 1797298410033774592 |
|---|---|
| author | Ruimin Chen Dailin Lv Li Dai Liming Jin Zhiyu Xiang |
| author_facet | Ruimin Chen Dailin Lv Li Dai Liming Jin Zhiyu Xiang |
| author_sort | Ruimin Chen |
| collection | DOAJ |
| description | Recent object detection networks suffer from performance degradation when training data and test data are distinct in image styles and content distributions. In this paper, we propose a domain adaptive method, Adversarial Mixing (AdvMix), where the label-rich source domain and unlabeled target domain are jointly trained by the adversarial feature alignment and a self-training strategy. To diminish the style gap, we design the Adversarial Gradient Reversal Layer (AdvGRL), containing a global-level domain discriminator to align the domain features by gradient reversal, and an adversarial weight mapping function to enhance the stability of domain-invariant features by hard example mining. To eliminate the content gap, we introduce a region mixing self-supervised training strategy where a region of the target image with the highest confidence is selected to merge with the source image, and the synthesis image is self-supervised by the consistency loss. To improve the reliability of self-training, we propose a strict confidence metric combining both object and bounding box uncertainty. Extensive experiments conducted on three benchmarks demonstrate that AdvMix achieves prominent performance in terms of detection accuracy, surpassing existing domain adaptive methods by nearly 5% mAP. |
| first_indexed | 2024-03-07T22:34:28Z |
| format | Article |
| id | doaj.art-53eabdf3bb6a4548af6504d507bfa026 |
| institution | Directory Open Access Journal |
| issn | 2079-9292 |
| language | English |
| last_indexed | 2024-03-07T22:34:28Z |
| publishDate | 2024-02-01 |
| publisher | MDPI AG |
| record_format | Article |
| series | Electronics |
| spelling | doaj.art-53eabdf3bb6a4548af6504d507bfa0262024-02-23T15:14:40ZengMDPI AGElectronics2079-92922024-02-0113468510.3390/electronics13040685AdvMix: Adversarial Mixing Strategy for Unsupervised Domain Adaptive Object DetectionRuimin Chen0Dailin Lv1Li Dai2Liming Jin3Zhiyu Xiang4Zhejiang Geely Holding Group Co., Ltd., Hangzhou 310051, ChinaZhejiang Geely Holding Group Co., Ltd., Hangzhou 310051, ChinaZhejiang Geely Holding Group Co., Ltd., Hangzhou 310051, ChinaZhejiang Geely Holding Group Co., Ltd., Hangzhou 310051, ChinaCollege of Information Science and Electronic Engineering, Zhejiang University, Hangzhou 310027, ChinaRecent object detection networks suffer from performance degradation when training data and test data are distinct in image styles and content distributions. In this paper, we propose a domain adaptive method, Adversarial Mixing (AdvMix), where the label-rich source domain and unlabeled target domain are jointly trained by the adversarial feature alignment and a self-training strategy. To diminish the style gap, we design the Adversarial Gradient Reversal Layer (AdvGRL), containing a global-level domain discriminator to align the domain features by gradient reversal, and an adversarial weight mapping function to enhance the stability of domain-invariant features by hard example mining. To eliminate the content gap, we introduce a region mixing self-supervised training strategy where a region of the target image with the highest confidence is selected to merge with the source image, and the synthesis image is self-supervised by the consistency loss. To improve the reliability of self-training, we propose a strict confidence metric combining both object and bounding box uncertainty. Extensive experiments conducted on three benchmarks demonstrate that AdvMix achieves prominent performance in terms of detection accuracy, surpassing existing domain adaptive methods by nearly 5% mAP.https://www.mdpi.com/2079-9292/13/4/685object detectiondomain adaptionadversarial learningself-training |
| spellingShingle | Ruimin Chen Dailin Lv Li Dai Liming Jin Zhiyu Xiang AdvMix: Adversarial Mixing Strategy for Unsupervised Domain Adaptive Object Detection Electronics object detection domain adaption adversarial learning self-training |
| title | AdvMix: Adversarial Mixing Strategy for Unsupervised Domain Adaptive Object Detection |
| title_full | AdvMix: Adversarial Mixing Strategy for Unsupervised Domain Adaptive Object Detection |
| title_fullStr | AdvMix: Adversarial Mixing Strategy for Unsupervised Domain Adaptive Object Detection |
| title_full_unstemmed | AdvMix: Adversarial Mixing Strategy for Unsupervised Domain Adaptive Object Detection |
| title_short | AdvMix: Adversarial Mixing Strategy for Unsupervised Domain Adaptive Object Detection |
| title_sort | advmix adversarial mixing strategy for unsupervised domain adaptive object detection |
| topic | object detection domain adaption adversarial learning self-training |
| url | https://www.mdpi.com/2079-9292/13/4/685 |
| work_keys_str_mv | AT ruiminchen advmixadversarialmixingstrategyforunsuperviseddomainadaptiveobjectdetection AT dailinlv advmixadversarialmixingstrategyforunsuperviseddomainadaptiveobjectdetection AT lidai advmixadversarialmixingstrategyforunsuperviseddomainadaptiveobjectdetection AT limingjin advmixadversarialmixingstrategyforunsuperviseddomainadaptiveobjectdetection AT zhiyuxiang advmixadversarialmixingstrategyforunsuperviseddomainadaptiveobjectdetection |