A scene information integrating network for object detection in remote sensing images

Object detection in remote sensing images is a challenging task in field of computer vision since detection performance is negatively influenced by complicated background and various object size. However, most studies have only focused on object appearance, with only a few taking into account scene information, which is closely related to existence and category of objects. In this paper, we put forward a new method by integrating scene information into detection with aim of generating more powerful feature. Specifically, we made use of GRU cell, a special kind of RNN, in order to enhance object feature. The proposed method was verified through experiments on a challenging dataset, i.e., DOTA. Compared to the baseline model RoI-Transformer, the proposed method has achieved around 2.7% improvement in terms of mAP, which is initial attempt to integrate scene information into object detection.