YOLOv5 for Enhanced Small Object Detection in Paired IR and Depth Images

This paper introduces a comprehensive approach to enhance Time-of-Flight (ToF) infrared image object detection. A novel Depth Images Enhancement method using Joint Filtering and Partial Convolution is proposed, simulating real-world distortions in low-quality depth maps. The Joint Depth Filtering Network and Partial Convolution are integrated to mitigate noise and invalid pixels. Additionally, the research refines the loss function selection for YOLOv5 in ToF image object detection. The adaptation of the Complete Intersection over Union Loss to Alpha Intersection over Union Loss ($\alpha=3$) enhances model robustness without introducing complexity. The refined loss function is expressed and validated, contributing to improved YOLOv5 performance. Experimental results demonstrate the effectiveness of the proposed algorithm.