SDCDet: Robust Remote Sensing Object Detection Based on Instance Segmentation Direction Correction

Aerial and satellite images usually contain complex background interference and dense small objects. The difficulties of remote sensing target detection tasks include complex scenes, large amounts of data, and the arbitrariness of target directions. Differing from the dominant regression-based approaches for orientation estimation, this paper combines example segmentation methods to design a dense small target detection and segmentation model. After passing through the feature extraction module, our method is divided into three branches: (1) Multiple Box Offset Regression Branch: used to perform dense small object frame regression; (2) Discriminative Classification Branch: can obtain more accurate classification results; (3) Segmentation correction branch: used to obtain instance segmentation output and correct the final Bounding box. We conducted training and testing on a large high-resolution remote sensing dataset (DOTA). Compared with the benchmark model, our model can obtain more accurate results.