Margin and Average Precision Loss Calibration for Long-Tail Object Detection

Large-scale object detection and instance segmentation face a severe data imbalance, seriously degrading their ability to detect rarely seen categories. Most existing approaches for long-tailed object detection rely on heuristic or experimental discovery, lacking theoretical foundations. They primarily focus on improving the classification scores of rare categories in the long-tailed data distribution, neglecting the impact of the long-tailed distribution on the detector's model average precision (AP). In this study, we introduce the theory of the most compact boundary of AP and propose a margin calibration loss. We integrate a feedback factor to dynamically adjust the model's classification boundaries. Moreover, by incorporating the AP loss, we investigate the impact of long-tailed object detection on AP computation and devise a calibration method to rectify biases induced by imbalanced data distribution. Experimental results substantiate the efficacy of our approach in augmenting the AP performance of detectors under long-tailed data distributions.