An Improved Mean Teacher Framework for Robust Domain Adaptive Foggy Object Detection

Abstract

To further address the limitations prevalent in existing domain adaptation methods for foggy object detection — namely, insufficient robustness against noise and interference, slow convergence during training, susceptibility to gradient oscillations and fluctuations, training instability, and inadequate extraction and utilization of small-object features — we propose a novel algorithm termed An Improved Mean Teacher Framework for Robust Domain Adaptive Foggy Object Detection (IMT-Det). First, the Mean Teacher paradigm is incorporated to alleviate the convergence difficulties inherent in adversarial training for domain-adaptive foggy object detection. The framework comprises a teacher model and a student model. During training, a dynamic smoothing coefficient is devised, through which the teacher model's parameters are updated via Exponential Moving Average (EMA) of the student model's parameters. This smooth training strategy mitigates convergence instability during adversarial optimization and, compared to a single-model baseline, renders the EMA-updated teacher model considerably less sensitive to noise and perturbations, thereby avoiding the oscillatory behavior that may arise from directly optimizing adversarial objectives. Subsequently, a weighted teacher loss mechanism is introduced, which dynamically adjusts loss weights to achieve fine-grained refinement of the training process and to resolve the problem of excessive regression errors under foggy scene conditions. Finally, a FEUCB upsampling module is designed to address the challenges posed by low contrast in foggy images and significant variations in object scale, thereby enhancing detection capability in foggy scenarios.