An Effective and Head-Friendly Long-Tailed Object Detection Network with Improving Data Enhancement and Classifier
Huiyun Gong, Jingxiu Yao, Yanhong Liu · 2023
Supervised learning methods achieve good results in large amounts of data and balanced data scenarios. However, the real world often has unbalanced data, such as long-tail distribution data sets, where the number of head categories is far exceeds the number of tail categories. Because the amount of tail data is small, using traditional supervised training methods will cause the model to be biased towards the head category. Therefore, when optimizing model performance, we often consider preventing bias from causing difficulty in improving model performance. A common approach is to balance the performance of the head and tail categories to achieve the purpose of improving the performance of the tail category and the overall performance of the model. Such approaches make a complex trade-off between the head and tail categories. Based on the above observations, we propose a head category-friendly long-tail object detection network (EHFL), including a data enhancement module (DEM) and a classifier enhancement module (CEM), and propose a decoupled performance adjustment strategy (DAS) for training, which can decouple the optimization of head category and tail category performance. In DEM, we propose using label information to generate training data to improve some categories’ performance. In CEM, we propose a classifer enhancement loss function to improve overall performance. Experimental results show that it reaches 26.48% on the LVIS-v0.5 data set, improving overall performance without reducing the performance of the head category, which is better than state-of-the-art methods.