رضا صديقي پاشاكي

Object detection is a critical area in computer vision an‎d deep learning that has consistently attracted the attention of researchers. However, adverse weather conditions such as foggy scenes, reduced image clarity, an‎d visual distortions caused by fog pose significant challenges to accurate object detection. Existing methods are generally categorized into three groups: sequential approaches, multi-task learning approaches, an‎d unsupervised domain adaptation methods, the latter of which has become the dominant strategy for tackling this challenge. Nevertheless, their reliance on labeled source data limits their applicability in privacy-sensitive scenarios. This study focuses on source-free object detection (SFOD). The chosen approach is the Mean Teacher framework, which has recently gained attention in self-training. However, it suffers from two fundamental issues: first, uncontrolled degradation of the teacher due to unstable updat‎es from the student; second, the student model’s tendency to replicate errors caused by incorrect pseudo-labels, trapping it in a local optimum. To overcome these issues, the referenced paper introduces a novel "Dynamic Retraining-Updating (DRU)" mechanism, which actively manages the studentʹs training process an‎d the teacher modelʹs updat‎es in a feedback-driven manner, ensuring their simultaneous evolution. eva‎luations on stan‎dard datasets demonstrate the superiority of this method over existing approaches an‎d confirm its generalization capability an‎d robustness across diverse environmental conditions. Among the practical contributions of this research is the successful application of gradient accumulation, which, by increasing the effective batch size, enabled leveraging the collective wisdom of data across multiple mini-batches, a strategy that facilitates training large models despite limited GPU memory. Additionally, the idea of strategically shifting the layer normalization position was implemented to improve model convergence an‎d performance. Although this approach did not achieve complete success due to challenges such as suboptimal hyperparameter tuning, it shows promising potential for future development. In conclusion, this thesis presents a collection of practical findings an‎d valuable suggestions that pave the way for future research in object detection under adverse weather conditions.

تشخيص اشياء بدون‌منبع (SFOD) , يادگيري عميق , معلم ميانگين , يادگيري خودآموز

Source-Free Object Detection (SFOD) , Deep Learning , Mean Teacher , Self-training

Reza Sedighi Pashaki

Dr. Sattar Mirzakuchaki