Dynamic Mutual Learning for Object Detection in Aerial Imagery

Object detection in aerial imagery is a pivotal task for various Earth observation systems, composed of two separate yet interdependent subtasks: classification and localization. However, existing methods face two fundamental limitations: 1) inconsistency in prediction distributions, where these two subtasks lack spatial distribution alignment, and 2) impracticability of cross-scale representations, where fixed-scale representations impede representation capacity and accuracy for objects of varying sizes in aerial scenarios. To overcome these challenges, this paper proposes a novel dynamic mutual learning paradigm that synergizes representation-wise and supervision-wise interactions within a unified detection head. It consists of two learning schemes: 1) dynamic learning, which introduces the dynamic routing mechanism to enable cross-scale fine-grained representation aggregation, significantly benefiting representational efficiency and flexibility, and 2) mutual learning, which establishes prediction alignment by explicitly performing subtask-consistent supervision and collaborative optimization. Moreover, within the entire enhanced detection head, these schemes can be jointly optimized and mutually reinforced. Extensive experimental results on different datasets have demonstrated the effectiveness and superiority of this proposed learning paradigm for object detection in aerial imagery, achieving competitive performance in both detection accuracy and computational efficiency.