三维点云场景识别与深度学习.pdfVIP

面向大规模场景识别和环境分析的 三维点云深度学习研究 LPD- Series ： ➢3D Point Cloud Learning for Large - Scale Place Recognition ➢Sequence Loop Closure Detection for Self - driving Vehicles 索传哲 Introduction to Large-scale Scene Research 1 Large-scale Place Description Series 2 C o n t e n t s Place Recognition Application cases 3 3D Point cloud Processing and Deep Learning 4 0 1 3D Point Cloud World Starfield Library in COEX Mall Seoul, Korea ➢ 3D data representation format: ➢ Academic community: very active from 2015: Large 3D dataset: • RGB-D image ShapeNet (Stanford), ModelNet (Princeton) • Mesh ➢ Industry community: broad applications • PointCloud • Robotics • Voxel • Autonomous driving • Virtual reality Xie, J. (n.d.). Deep Learning Based 3D Shape Representation. ➢ Challenges in 3D deep learning: ➢ Large deformations of 3D shapes • 3D model: geometric structure information ; ➢ Large structure variations of 3Dshapes • 2D image: pixel value ➢ Partial models of 3D shapes • 3D model: irregular data structure; • 2D image: regular data structure • Unordered • Invariance under transformations Xie, J. (n.d.). Deep Learning Based 3D Shape Representation. ➢ Deep lear