A New Monocular 3D Object Detection with Neural Network
Coherently recognizing objects and localizing its 3D position is vital in a large number of practical applications such as robot,autonomous unmanned systems,etc.To tackle this problem,we present a new framework composed of two coupled deep networks.The framework involves three tight steps: first,by implanting a pyramid strategy into the end-to-end detection network,the goal of detecting and recognizing objects with large or small scales is attained in 2D pictures; second,by a novel lost function we proposed for the depth networks,the 3D points of the objects are accurately estimated; and finally,we propose a windowed point cloud segmentation: combing 3D norm direction,curvature and color information,outlier 3D points are removed from windows,thus the clean object is extracted with its 3D position.The obvious advantage of the framework lies on that most the steps are implemented by the fused networks,and this makes the frame-work efficient,accurate and robust provided with enough training data.The framework is evaluated on the KITTI benchmark,and the efficiency is superior to the state-of-the-art methods while the accuracy is comparable.Our method is several tens of times faster than most of previous methods and capable of applications with real time constraint.In addition,we only use weak supervision learning methods in our framework that need less manual annotations.
3D Object Detection Deep network Windowed point cloud segmentation
Weijie Hong Yiguang Liu Yunan Zheng Ying Wang Xuelei Shi
Vision and Image Processing Lab(VIPL),College of Computer Science,Sichuan University,Chengdu 610065,Peoples Republic of China
国际会议
广州
英文
174-185
2018-11-23（万方平台首次上网日期，不代表论文的发表时间）