Aligning Localization and Classification for Anchor-Free Object Detection in Aerial Imagery

Anchor-free aerial object detectors have recently attracted considerable attention due to their high flexibility and computational efficiency. They are typically implemented by learning two subtasks of object detection, object localization and classification, based on two separately parallel branches in the detection head. However, without the constraints of predefined anchor boxes, anchor-free detectors are more vulnerable to spatial misalignment caused by optimization inconsistencies between these two subtasks, which significantly degrades detection performance. To address this issue, this paper proposes a novel and efficient anchor-free object detector, namely localization-classification-aligned detector (LCA-Det), which explicitly pulls closer the predictions of localization and classification, through a single-branch subtask-aligned detection head and a subtask-aligned sample assignment metric. Extensive experimental results have demonstrated the effectiveness and superiority of our proposed method for object detection in aerial imagery.