A novel few-shot object detection framework for multi-scene driving based on contrastive proposal encoding

This paper proposes a few-shot object detection algorithm based on FSCE, tailored for multi-scenario driving environments. Unlike existing methods that focus on single scenarios, our approach addresses challenges of cross-scenario heterogeneity and overfitting in low-data regimes. We enhance feature representation through a multi-scale feature module that integrates local and contextual information, and replace the traditional Softmax with a cosine Softmax classifier to reduce intra-class variance via L2 normalization and angular margin constraints. This work is the first to apply few-shot detection to both nighttime infrared and daytime visible-light driving scenarios. Experiments on FLIR and BDD100K demonstrate superior generalization and accuracy over existing methods. Future work will explore reducing model complexity while maintaining performance.