A Weighted Function Based on Loss Distribution in Face Detection

Abstract In recent years, advancements in deep learning and general object detection technologies have significantly enhanced the performance of face detectors through optimizations in network architecture, loss functions, and other aspects. However, in anchor-based face detection methods, there remains the issue of uneven distribution of matched anchors for different faces. Existing methods often utilize data-based methods and compensate additional anchors for outer faces to enhance detector performance. However, these methods typically focus on a subset of faces, neglecting the optimization of each individual face during training. Our experiments investigate the distribution of classification and localization losses of faces matched with varying numbers of anchors. We propose a weighted function to alleviate the contradiction, enabling the detector to adjust losses on a per-face basis across various subtasks. This approach facilitates the exploration of hard samples and balances losses across different faces. Finally, our model is validated on the WIDER FACE and FDDB datasets, demonstrating a good balance between accuracy and computational overhead.