MIAYOLO: Multi-Expert and Intra-Class Aggregation Assisted Suburban Building Detection in Unmanned Aerial Vehicle Imagery

In recent years, the utilization of unmanned aerial vehicle (UAV) for building detection has become increasingly important in monitoring illegal construction. However, the UAV imagery contains a significantly larger number of completed buildings compared to under construction buildings, which leads to the model exhibiting a bias toward the recognition of completed buildings during the training. Moreover, numerous under construction buildings visually resemble completed buildings, posing a challenge for existing object detection models in distinguishing between these two categories. To address these issues, in this study, we propose multi-expert and intra-class aggregation assisted detection (MIAYOLO) for building detection in UAV images. Specifically, a multi-expert subnetwork is proposed to improve the learning of distinct feature representations for each category, which mitigates the model bias toward the dominant category. Furthermore, an intra-class loss is introduced to aggregate features from the same category while separating instances from the decision boundary. Additionally, the multi-expert subnetwork is discarded to avoid introducing additional computation in the inference stage. Experimental results on our proposed dataset demonstrate that MIAYOLO outperforms state-of-the-art object detection methods, achieving a significant improvement of +4.7 (in terms of mAP50). The code and dataset can available at https://github.com/Game-zhao/MIAYOLO.