SPL: stable pseudo-label learning for domain adaptive object detection
Domain-adaptive object detection uses labelled source images and unlabelled target images to address distribution shift. We introduce stable pseudo-label learning (SPL), an incremental extension of the JMT-YOLO Mean Teacher framework that couples adaptive filtering with dynamic assignment. Distribution-aware adaptive thresholding (DAAT) applies Otsu’s between-group variance criterion to a 100-bin histogram of teacher confidence scores and smooths the resulting threshold over iterations. Dynamic pseudo-label assignment (DPLA) restricts candidates by pseudo-box centres, combines classification, overlap and centre-distance costs, and uses IoU-based dynamic positive counts with minimum-cost conflict resolution. SPL obtained 48.6% and 58.4% [email protected] on Pascal VOC \(\rightarrow\) Clipart1k and Cityscapes \(\rightarrow\) Foggy Cityscapes, respectively, and 62.8% Car [email protected] on KITTI \(\rightarrow\) Cityscapes. The corresponding gains over JMT-YOLO were 1.7, 1.1 and 1.8 percentage points. Component ablations support an incremental detection benefit in the evaluated framework. The evaluation covers style, weather and camera shifts within JMT-YOLO, with target-validation-informed model selection.
Authors
- Wenchao Zhang (ORCID: https://orcid.org/0000-0003-1989-3032)
- Zhenhua Yu (ORCID: https://orcid.org/0000-0001-7548-3869)
- Yanlong Zhao
- Xuedi Hao
- Jiahao Zhang
Institutions
- Xi'an University of Science and Technology (CN)
- China University of Mining and Technology - Beijing (CN)
Publication Details
- Journal
- Scientific Reports
- Published
- 2026-11-14
- DOI
- https://doi.org/10.1038/s41598-026-78053-9
- Primary Topic
- Domain Adaptation and Few-Shot Learning
- Type
- article
- Field-Weighted Citation Impact
- 0.00
Funders
- Key Research and Development Projects of Shaanxi Province