Oriented Object Detection For Remote Sensing Images Based On Weakly Supervised Learning
Yongqing Sun, Jie Ran, Feng Yang, Chenqiang Gao, Takayuki Kurozumi, Hideaki Kimata, Ziqi Ye
Abstract
Object detection of remote sensing images (RSIs) is an active yet challenging task because of the complex appearance of ground objects and the particular imaging views. One of the difficulties in RSI object detection is the orientation variation, where the objects could take on arbitrary orientations due to the birdview shot from high altitudes. For oriented object detection, existing methods rely on largescale dense oriented annotations for training deep networks under full supervision, which are resource-intensive. To address this problem, (a) we propose a kind of weakly supervised oriented object detection method in this paper. With only the horizontal-object supervision, we rotate object proposals via an angle search strategy to align them as horizontally as possible and detect the oriented objects just like the horizontal ones. We aim to mine more oriented objects and thus can train the Rotational RCNN framework. Experimental results demonstrate that our method can achieve significant performance improvement on the oriented object detection and outperforms the state-of-the-art methods.