Exploring Classification Equilibrium in Long-Tailed Object Detection

"Exploring Classification Equilibrium in Long-Tailed Object Detection" is a research paper that discusses the problem of long-tailed distribution in object detection tasks. Long-tailed distribution means that there are many classes with very few samples, while a few classes have a large number of samples. The paper proposes a method called Class-Balanced Loss with Dynamic Scaled Sampling (CB-DSS) to address this problem. CB-DSS involves re-balancing the loss function and sampling strategy to give more importance to underrepresented classes. The paper also explores the concept of classification equilibrium, where the classifier is trained to maintain a balance between the accuracy of rare and frequent classes. The proposed method is evaluated on multiple datasets and is shown to outperform existing state-of-the-art methods in long-tailed object detection tasks.