D2ETR: A Decoupled DETR for Efficient Detection in Aerial Images
Changfeng Feng, Chunping Wang, Qiang Fu, Renke Kou · 2024
Objects in aerial images possess distinctive features such as large-scale variations, intricate backgrounds full of distractions, and versatile viewpoints. These factors present significant challenges for common object detectors based on convolutional networks. The latest transformer-based detector, DEtection TRansformer (DETR), performs impressively. Nevertheless, the attention mechanism's numerous operations limit its application to high-resolution aerial imagery. We have analyzed the design of the DETR-like detector and put forward Decoupled DETR (D2ETR), which aims to process multiscale feature information of aerial images efficiently to balance computational efficiency and accuracy. Our proposal involves an effective decoupled encoder that handles long multiscale feature sequences. The encoder comprises two modules: an attention-based semantic enhancement module and a convolution-based cross-scale fusion module. We have developed a feature stabilization module to counter feature information disorder caused by different processing mechanisms. In addition, we have applied a small-object-friendly loss function to the prediction layer to improve the model's ability to adapt to small targets in aerial images. Experimental results using the VisDrone and DIOR datasets demonstrate our approach reduces computation while maintaining the transformer's excellent performance.