Model-Independent Approach For Long-Tail Object Detection In Aerial Imagery

The class imbalance and long-tail characteristics of real-world aerial datasets make object detection methods difficult to improve their performance. Disparities in class distribution, particularly between long-tail and head classes (foreground-foreground disparity), impede the efficacy of less prevalent categories. Thus, the goal of this work is to improve the performance of tail categories in long-tail aerial datasets while enhancing or at-least maintaining the performance of head categories. To achieve this, a weighting and re-weighting strategy is proposed that focuses on the foreground-foreground disparity in classes at the dataset level, thereby enhancing the accuracy of less prevalent categories. The proposed strategy is evaluated using two long-tail aerial datasets, the drone-based VisDrone DETection (VisDroneDET) and the Satellite Imagery Multivehicles Dataset (SIMD), and one-stage object detection models. Compared to baseline models, most tail categories show significant performance improvements, highlighting the importance of the weighting strategy to address class imbalances in long-tail object detection.