Foreground and Background Separate Adaptive Equilibrium Gradients Loss for Long-Tail Object Detection

Tianran Hao, Ying Tao, Meng Li, Xiao Ma, Peng Dong, Lisha Cui, Pei Lv, Mingliang Xu

Abstract: The current mainstream object detection methods usually tend to implement on datasets where the categories remain balanced, and have made great progress. However, in the presence of long-tail distribution, the performance is still unsatisfactory. Long-tail data distribution means that a few head classes occupy most of the data, while most of the tail classes are not representative, and tail classes are excessive negatively suppressed during training. Existing methods mainly consider suppression from negative samples of the tail classes to improve the detection performance of the tail classes, while ignoring suppression from correct background prediction. In this paper, we propose a new Foreground and Background Separate Adaptive Equilibrium Gradients Loss for Long-Tail Object Detection (FBS-AEGL) to deal with the problem mentioned above. Firstly, we introduce the numerical factor among categories to weight different classes, then adaptively leverage the suppression of head classes according to the logit value of the network output. Meanwhile, dynamically adjusting the suppression gradient of the background classes to protect the head and common classes while improving the detection performance of the tail classes. We conduct comprehensive experiments on the challenging LVIS benchmark. FBS-AEGL Loss achieved the competitive results, with 29.8% segmentation AP and 29.4% box AP on LVIS v0.5 and 28.8% segmentation AP and 29.4% box AP on LVIS v1.0 based on ResNet-101.