【泡泡机器人公开课预告】苏昊-Deep Part Induction from 3D Shapes

2018年8月15日 （周三）上午 10:00，由苏昊老师给大家带来一次公开课，公开课题目为：Deep Part Induction from 3D Shapes（3D形状的深层感应）。

苏昊，加州大学圣地亚哥分校计算机科学与工程专业的助理教授，他隶属于语境机器人研究所和视觉计算中心。他的实验室（SU Lab）致力于结构理解、形状理解和与人工智能相关的场景理解问题。他现在领导着ShapeNet的建设，这是一个大规模的以3d为中心的物体（SGP数据集）的知识库，并曾参与了ImageNet数据集的相关工作，这是一个大型的2D图像数据库（PAMI Mark Everingham奖）。

 他曾在计算机视觉、计算机图形学和机器学习等多个会议和研讨会的项目委员会任职。他是2019年CVPR的领域主席，2018国际环太平洋图形会议，2017年3DV会议的项目主席，2016年3DV会议的出版主席，以及CVPR，ECCV和ICCV的多个研讨会的主席。在NIPS、3DV、CVPR和S3PM等研讨会和专题报告中，他还被邀请作为主题演讲嘉宾。根据谷歌学术的记录，他的研究已经收到了超过1万次的引用。

原文：

Hao Su is an Assistant Professor of Computer Science and Engineering atUCSD. He is affiliated with the Contextual Robotics Institute and Center forVisual Computing. His lab (SU Lab) works on Structure Understanding, ShapeUnderstanding, and Scene Understanding problems that are relevant to artificialintelligence. He is now leading the construction of ShapeNet,a large-scale 3D-centric knowledge base of objects (SGP Dataset Award), andused to work on ImageNet,a large-scale 2D image database (PAMI Mark Everingham Prize).

He served on the program committee of multiple conferences andworkshops on computer vision, computer graphics, and machine learning. He isthe Area Chair of CVPR'19, IPC of Pacific Graphics'18, Program Chair of 3DV'17,Publication Chair of 3DV'16, and chair of various workshops at CVPR, ECCV, andICCV. He is also invited as keynote speakers at workshops and tutorials inNIPS, 3DV and CVPR, S3PM, etc. His research has received over 10,000 citations,according to Google Scholar.

时       间

北京时间2018年8月15日 10:00

 内       容

理解3D形状的主要目标之一是提取对象结构，例如，将对象分解成多个部分。近年来，随着大型3D数据集的使用和深度学习技术的应用，有监督分割已经取得了显著的进展。尽管这一进展令人印象深刻，但当涉及到在训练集中没有出现过的新对象类别的3D形状时，现有算法的能力仍远不及人类。能够解析没有见过的新形状对构建智能代理至关重要,包括在增强现实场景、虚拟世界中帮助人类或在自主机器人设置中服务人类等。因为随着新对象不断出现,代理必须能够从少量的观察和交互体验中提取知识来建模和适应新环境。

在这节课中，我将介绍我最近在结构提取上的工作，将深度学习与其他机器学习技术结合起来，以解决不同环境下的挑战，例如完全无监督（CVPR'17的工作），弱监督（arxiv）和few-shot learning（小样本学习）（SIGGRAPH Asia 2018）。

原文：

One of the primary goals of understanding 3D shapes is to extractobject structures, e.g., to decompose objects into parts. Recently, with theavailability of large 3D datasets and the use of deep learning techniques,significant progress has been made in the task of supervised part segmentation.Although the progress is impressive, however, contemporary algorithms are stillfar inferior to humans when it comes to parsing 3D shapes from novel objectcategories that are not covered in the training sets. Being able to parse newshape categories not seen before is fundamentally crucial towards buildingintelligent agents that can assist humans in augmented reality scenarios,virtual worlds, or serve humans in autonomous robotics settings: as new objectscontinuously emerge, an agent must model and adapt to novel environments byinducing knowledge from few observations and interaction experiences.

In this lecture, I willintroduce my recent efforts on the part structure discovery problem bycombining deep learning with other machine learning techniques to address the challengesin different settings, such as purely unsupervised (CVPR’17 work), weaklysupervised (arxiv), and few-shot learning (SIGGRAPH Asia 2018) settings. 

参与方式

此次分享采用斗鱼直播，长按下图，识别图中二维码，关注“泡泡机器人SLAM”（ID：paopaorobot_slam）公众号，回复 "公开课" 获取直播链接，提前关注直播间，我们将在上午10:00开始直播，欢迎大家参加分享！

欢迎大家联系泡泡助手（ID:paopaorobot_service)加入泡泡粉丝群一起讨论

编辑：王琛

 ：：END：：