Accurate Bridge Detection in Aerial Images With an Auxiliary Waterbody Extraction Task

Bridge detection in aerial images is to determine whether a given image contains one or more bridges and localize them. However, the arbitrary orientations, extreme aspect ratios, complex backgrounds pose great challenges for bridge positioning. In this paper, we tackle these problems by combining strengths of semantic-segmentation-based auxiliary supervision, waterbody constraint, instance-switching-based data augmentation. More precisely, make three main contributions: (i) We propose an oriented model with task segmentation, which performs as guidance localization. The network specifically designed cascade style handle segmentation end-to-end. (ii) use semantic features spatial attention distinguish from cluttered backgrounds, then generate map introduces prior knowledge distribution refine predictions. (iii) background consistent instance switching method online augmentation further improve robustness detection. To verify effectiveness proposed method, introduce dataset named BridgeDetV1 containing 5,000 well-annotated two representations, i.e., horizontal bounding box box. Extensive experiments demonstrate that our approach outperforms state-of-the-art methods on challenging benchmark. Dataset, code, models are available at https://github.com/whughw/BridgeDet.