Masked Image Self-Learning and Knowledge Distillation for Source-Free Unsupervised Domain Adaptation in Crowd Counting
Jia Zeng, Chaoqun Ma, Penghui Shao, Anyong Qing, Yang Wang · 2023
Unsupervised Domain Adaptation (UDA) techniques leverage labeled data from a source domain to adapt to unlabeled data from a target domain, offering a primary solution for addressing cross-domain challenges in crowd counting. However, in real-world scenarios, accessing source data during the model adaptation process within the target domain can be impractical due to privacy and security concerns. This gives rise to a practical yet challenging issue known as Source-Free Unsupervised Domain Adaptation (SFUDA), where model adaptation is limited to the utilization of an existing source model and unlabeled target data. To tackle this challenge, we propose an algorithm that combines mask image self-learning and knowledge distillation to address SFUDA in crowd counting effectively. Specifically, the student model uses masked images to self-learn at both feature and output levels, exploring clues and contextual relationships around masked image blocks to predict occluded areas. Furthermore, we employ knowledge distillation to boost the model’s performance within the teacher-student framework, encompassing feature extraction and density map generation stages. To mitigate cumulative errors, we stochastically recovery a subset of parameters from the source model. We conducted extensive experiments to compare our approach with a modified SFUDA algorithm, and the results demonstrate competitive performance across multiple datasets.