Object Detection Transfer Learning paradigm on Cross Domain Street Vehicle Detection
Shu Fang Zhang, S. Hsu, Chi Han Chen, Rung‐Shiang Cheng · 2024
In street vehicle detection, varying environmental and urban conditions necessitate extensive data collection for model training. When adapting a high-precision model to a new urban area, it is crucial to integrate data from this new domain to prevent overfitting to prior knowledge, which could impair performance in new tasks. Transfer learning, specifically through knowledge distillation and cross-domain adaptation, utilizes knowledge from a pre-existing model. Our approach introduces a training method that incorporates minimal data from the previous context into the new task and applies fine-tuning for cross-domain learning. Our experiments demonstrate that this method, augmented with source-approximate data, significantly enhances adaptability in target domains for vehicle detection, outperforming models based solely on knowledge distillation.