Class Distribution Aware Focal Loss for Long-Tailed Object Detection

Object detection tasks have made significant progress on class-balanced, high-quality datasets. However, training data often exhibit a long-tail distribution in real-world scenarios. Existing re-weighting methods have made preliminary explorations into the imbalance issue of positive and negative samples across different categories. However, the additional hyperparameters introduced complicate the network’s practicality and performance on the overall dataset. To address this, we propose a class-distribution-aware focal (CDAF) loss. Without the need for meticulous manual parameter tuning, it can dynamically control the contribution of different sample categories and complexity levels to model training using a balancing factor related to sample features. This allows the model to improve performance not only on rare classes but also on difficult samples. Extensive experimental results demonstrate that our approach achieves significant performance improvements on challenging object detection datasets, including Pascal VOC and MS COCO.