3DV 2020 Tutorial on

3D Point Cloud Reconstruction and Segmentation

Anywhere on earth on November 28, 2020
Tutorial videos and slides will be released after the online events.



With the rapid development of point-cloud acquisition techniques (e.g., Microsoft Kinect) and computing devices, 3D data (point clouds, depth images, meshes) processing has become a rapidly growing research area in computer vision, pattern recognition and computer graphics. Extensive investigations have been conducted on 3D related topics, such as 3D modeling, 3D scene reconstruction/understanding, 3D object recognition, 3D face recognition, and 3D animation. In this tutorial, we will mainly focus on point cloud reconstruction and segmentation.

Existing 3D point cloud segmentation methods can be broadly divided into two categories based on the segmentation granularity, i.e., semantic segmentation and instance segmentation methods. Due to the unique properties (i.e., unstructured, irregular, and orderless) of 3D point clouds, it is still highly challenging to achieve fast and robust segmentation of 3D point clouds. Recently, a large number of 3D point cloud segmentation algorithms have been proposed in literature. The proposed tutorial will therefore present a comprehensive review and analysis of the state-of-the-art 3D segmentation algorithms. The tutorial will also provide extensive performance evaluation results of the state-of-the-art algorithms on several benchmark datasets, along with insightful discussions and analyses. Moreover, a number of interesting 3D related applications will be introduced in the tutorial. Finally, several directions for future work will be discussed.

The main objective of the tutorial is to stimulate communication between researchers from different areas (including computer vision, computer graphics, and machine learning) and from different sectors (e.g., academia and industry). Particularly, this tutorial will bridge the gap between different communities by presenting most existing 3D point cloud reconstruction and segmentation pipelines in a unified framework.


Time (GMT) Time (EST) Time (Beijing) Time (Tokyo) Topic
08:00-08:05 04:00-04:05 16:00-16:05 17:00-17:05 Welcome and Introduction
08:05-08:50 04:05-04:50 16:00-16:05 17:05-17:50 Background & Applications
Presenter: Dr. Yulan Guo
  • 3D data acquisition
  • 3D data representation
  • Point cloud learning
  • Applications
08:50-09:35 04:50-05:35 16:50-17:35 17:50-18:35 3D Point Cloud Reconstruction
Presenter: Dr. Ronald Clark
  • Dense point clouds as a map representation
  • Basics of SLAM and multi-view 3D reconstruction
  • State-of-the-art methods
  • Open challenges and new directions
09:35-10:20 05:35-06:20 17:35-18:20 18:35-19:20 3D Point Cloud Semantic Segmentation
Presenter: Mr. Qingyong Hu
  • Main challenges of 3D point cloud semantic segmentation
  • How to perform 3D point cloud semantic segmentation
  • Taxonomy: discretization, point and hybrid based algorithms
10:20-11:05 06:20-07:05 18:20-19:05 19:20-20:05 3D Point Cloud Instance Segmentation
Presenter: Dr. Bo Yang
  • Why and what is 3D point cloud instance segmentation
  • How to perform 3D point cloud instance segmentation
  • Taxonomy: proposal-based and proposal free algorithms
  • State-of-the-art algorithms
  • Performance evaluation
  • Summary and discussion
11:05-11:50 07:05-07:50 19:05-19:50 20:05-20:50 Panel Discussion
Host: Dr. Yulan Guo
  • Panelist introduction
  • Panelist presentations
  • Moderator-curated questions directed to the panelists
  • Questions from the audiences directed to the panelists
  • Virtue coffee break
11:50-12:00 07:50-08:00 19:50-20:00 20:50-21:00 Acknowledgments
*Please click here to know hours in your time-zone.

Please contact Qingyong Hu if you have question.