Cosine Similarity Guided Joint Modeling of Foreground and Background for Incremental Object Detection in Drone Images

Incremental object detection (IOD) aims to learn the localization and classification of new classes from continuously arriving data, while maintaining performance on previously learned old classes. Existing research to incremental object detection primarily focuses on natural scenes; however, the dynamic characteristics of aerial scenarios present unique challenges. These scenarios often involve small, numerous objects that can easily be confused with the background, resulting in inadequate performance for incremental object detection in drone applications. In our study, we investigate incremental object detection specifically in aerial scenarios. We have identified that existing incremental object detection methods exacerbate the confusion between old classes and the background when applied to drone environments. To address this issue, we propose a foreground and background joint modeling approach based on cosine similarity. First, on the basis of standard knowledge distillation method, we utilize old-class model to extract old-class foreground, which supplements the missing annotations for old classes in incremental steps. Second, during calculation of classification loss, we incorporate joint modeling of old classes and the background class. To obtain the probability prediction of negative samples, besides the background probability output, the old categories with the highest normalized logits are also considered. Their probability predictions are added together for the calculation of classification loss, utimately utilized for gradient backpropagation of background class. We conduct extensive experiments on the VisDrone dataset, which consists of aerial images. Our method demonstrates significant improvements over current state-of-the-art baseline methods under various incremental settings, with the Average Precision (AP) improvement by 0.64% and 0.73% in the two-step and multi-step settings, respectively.