No-Extra Components Density Map Cropping Guided Object Detection in Aerial Images

Aerial images usually contain a large number of truncated and small objects, which poses a significant challenge for object detection. Existing methods have introduced additional learnable components in the pipeline and adopted multistage training approaches, but they have not solved the problem of achieving end-to-end detection. To address this issue, we propose a novel no-extra components density map cropping (NE-CDMNet) method to utilize the spatial and contextual information between objects to improve detection performance. Furthermore, we introduce a new query selection (QS) scheme that utilizes confidence scores to select the top-K features from the encoder, helping the model better leverage the position information for extracting more comprehensive content features. Finally, we incorporate the local-global fusion (LGF) algorithm to combine the detection results from the original image and the density-cropped image. We conducted extensive experiments on two widely used public aerial datasets. Results reveal that the proposed method achieves the best performance compared with other state-of-the-art methods, whose 35.3% AP on the VisDrone-DET2019 dataset and 79.7% mAP on object detection in optical remote sensing image (DIOR) dataset, demonstrate the effectiveness of our method.