Abstract
Object detection in remote sensing images is challenging due to the dense distribution and arbitrary angle of the objects. It is a consensus that the oriented bounding box (OBB) is more suitable to represent the aerial objects. However, there are some extreme cases in regression-based OBB detection that make the regression target discontinuous, resulting in the poor performance. In this article, an analysis of the formats of OBB and the problems in its regression is presented, following with an exploration of transform localization from regression to keypoint estimation, which could be applied to avoid the problem of discontinuous regression target. Our novel method is called Object-wise Point-guided Localization Detector (OPLD). Continuously, a new prediction of center-point is introduced to refine the results, as the truncation problem caused by the cut graph. Lastly, in order to figure the problem of inconsistency between the localization quality and the classification score, both the endpoint scores and the classification score are adopted weighting as a result score. Experimental results are based on two widely used datasets, i.e., DOTA and HRSC2016. OPLD achieve 76.43% mAP and 78.35% mAP in OBB and horizontal bounding boxes tasks of DOTA-v1.0, which achieves state-of-the-art performance, respectively. Project page at https://github.com/yf19970118/OPLD-Pytorch.
Highlights
O BJECT detection is an essential task in computer vision, which can be decoupled into object classification and location
Our method uses the four endpoints of oriented bounding box (OBB) as key points, and the proposals generated by region proposal network (RPN) ensure that every four endpoints directly correspond to a particular object, avoiding the problem of grouping
1) DOTA: We compare Object-wise Point-guided Localization Detector (OPLD) with the state-of-the-art methods on OBB and horizontal bounding box (HBB) tasks of DOTA dataset in Tables V and VI
Summary
O BJECT detection is an essential task in computer vision, which can be decoupled into object classification and location. The keypoint-based method [5]–[7] detects all points on the entire image used to represent the bounding box and groups them to obtain the final result. Many works [9]–[12] have explored the use of oriented bounding box (OBB) and have made significant progress Among these methods, regression-based detection is the mainstream. With the help of region proposal network (RPN), we first obtain the horizontal circumscribed rectangle of each object, classify and detect the keypoints inside the proposal At this time, each endpoint corresponds to an object uniquely, avoiding point combination in the mainstream keypoint-based detector. By converting the original offset regression into unique point-guided keypoint estimation for localization, our method avoids discontinuous regression targets and achieves the state-of-the-art performance. By converting the original offset regression into unique point-guided keypoint estimation for localization, our method avoids discontinuous regression targets and achieves the state-of-the-art performance. 2) We add the center point prediction in the network and use it for center-point postprocessing (CPP) to solve the endpoint loss problem caused by remote sensing image cropping. 3) We use the average localization score of the OBB endpoints to correct the classification score to obtain the final score, which improves the correlation between classification confidence and localization quality
Sign-in/Register to view highlights & summary Sign-in/Register to access full text options 
Free) 
Talk to us
Join us for a 30 min session where you can share your feedback and ask us any queries you have
Schedule a call
