Enhancing Cross-Domain Vehicle Detection with Transfer Learning and Source-Similar Sample Integration

Abstract

In practical street vehicle detection applications, models may require a large amount of data due to varying street conditions across different regions, influenced by factors such as shooting angles and weather changes. Even with a high-precision detection model, applying it to a new urban area requires incorporating data from the new domain into the training process. To leverage the knowledge of an existing model for a new task, transfer learning models are utilized to prevent over-reliance on previous knowledge during training, which might result in the inability to detect target samples in the new task. Common research and application methods include knowledge distillation and cross-domain adaptation. This paper introduces a training paradigm that involves incorporating a small amount of source-approximate samples from the old task into the new task, followed by fine-tuning to experiment with cross-domain learning applications. Experimental results demonstrate that our paradigm, when augmented with source-approximate data—samples with similar scene or weather characteristics to the source domain—exhibits higher adaptability for detecting vehicle objects in the target domain compared to models utilizing knowledge distillation.