ROPDet: real-time anchor-free detector based on point set representation for rotating object
- PDF / 2,090,735 Bytes
- 12 Pages / 595.276 x 790.866 pts Page_size
- 38 Downloads / 158 Views
SPECIAL ISSUE PAPER
ROPDet: real‑time anchor‑free detector based on point set representation for rotating object Zhixiang Yang1 · Kunkun He4 · Fuhao Zou2 · Wanhua Cao1 · Xiaoyun Jia3 · Kai Li4 · Chuntao Jiang5 Received: 25 March 2020 / Accepted: 24 August 2020 © Springer-Verlag GmbH Germany, part of Springer Nature 2020
Abstract Remote-sensing object detection is a challenging task due to the difficulties of separating the objects with arbitrary direction from complex backgrounds. Though substantial progress has been made, there still exist challenges for object detection under the scenario of small scale, large aspect ratio, and dense distribution. Besides, the current mainstream approach falls under anchor-based multi-stage method, which has a serious shortcoming of slower inference speed. To conquer the aforementioned issues, this paper used RoPoints (points in rotation objects), a new better representation of objects as a set of sample points to perform object localization and classification. Then, we propose an anchor-free refined rotation detector:ROPDet based on RoPoints for more accurate and faster object detection. In our method, there is no need to predefine a large number anchors with different shapes. We only need to learn RoPoints for each object followed by converting to the corresponding bounding box, which greatly accelerates the inference process. Extensive experiments on two public remote-sensing datasets DOTA and HRSC-2016 demonstrate the competitive ability in terms of accuracy and inference speed. Keywords Remote-sensing image · Object detection · Arbitrary direction · RoPoints
1 Introduction Object detection is a fundamental and challenging task in computer vision field, which contains two main tasks: recognition and localization. It is widely used in the fields of automatic driving, intelligent video surveillance analysis and face detection, etc. With the rapid development of deep learning [23, 29, 35, 43], object detection has made a great progress. The mainstream object detection algorithms can be * Fuhao Zou [email protected]; [email protected] Zhixiang Yang [email protected] 1
Wuhan Digital Engineering Research Institute, Wuhan, China
2
School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China
3
School of Natural and Computational Science, Massey University, Wellington, New Zealand
4
School of Computer Science and Technology, Huazhong University of Science and Technology, Wuhan, China
5
School of Mathematics and Big Data, Foshan University, Foshan, China
divided into two categories: one-stage methods [25, 27, 31] and multi-stage methods [6, 12, 13, 22, 24, 32]. Generally, the one-stage methods have faster inference speed, while the multi-stage methods offer a better accuracy over the public benchmarks such as COCO [26] and VOC2007 [11], etc. In the real world, the majority of the object detectors fall under horizontal category which cannot be applicable to the scenario of remote-sensing image detection and
Data Loading...
