活动通知：GAMES Webinar339期-几何配准与生成（9月12日）

报告题目

Semantic Shape Registration 

Empowered by Diffusion-based 

Morphing and Flow Estimation

报告嘉宾

黄儒麒

清华大学深圳国际研究生院

报告时间

2024年9月12号 晚上8:00-8:30（北京时间）

报告方式

GAMES直播间: https://live.bilibili.com/h5/24617282

报告摘要

In this talk, I would like to introduce one of our recent works about shape analysis based on large vision models. In particular, we consider the problem of estimating dense correspondences between two shapes undergoing significant geometric variabilities. Prior purely geometric works mainly follows a general scheme of first estimating/annotating sparse landmark correspondences and then refining such for dense maps. In contrast, we tackle this problem by semantically and continuously morphing one shape to the other. More concretely, we first leverage diffusion-based image morphing and multi-view reconstruction techniques to establish a coarse morphing sequence between input shapes. Then, taking the former as auxiliaries, we estimate the regarding flow field to achieve interpolation and therefore dense correspondences. We demonstrate that our method achieves high-quality interpolation and maps between challenging shape pairs from a range of different categories, without any manual input or category-specific training.

黄儒麒是清华大学国际研究生院的助理教授。他在2011，2013年分别于清华大学数学科学系获的学士、硕士学位，之后在2016年于法国巴黎萨克雷大学计算机科学系获得博士学位。加入清华大学之前，他在法国巴黎综合理工学院计算机科学系任博士后研究员。他的研究工作主要围绕三维计算机视觉、几何数据处理展开。近年来他的工作在Science Advances，IEEE TPAMI， CVPR，ICCV，NeurIPS等顶级期刊、会议上获得发表。

https://rqhuang88.github.io

报告题目

Exploring Representations 

for 3D Generative Models

报告嘉宾

张彪

KAUST

报告时间

2024年9月12号 晚上8:30-9:00（北京时间）

报告方式

GAMES直播间: https://live.bilibili.com/h5/24617282

报告摘要

In this study, we investigate generation models for 3D shapes, representing objects as signed distance functions (SDFs). Firstly, we introduce sparsity into 3d object representations, demonstrating its utility as ga generative model and affirming the representation. Subsequently, we explore compressing objects entirely into a set of vectors, which not only incorporates sparsity but also elimnates the need for manually designing sparse structures. We establish its efficacy as a generative model, yielding exceptionally good generation results. Furthermore, we reveal that 3d object representation is essentially an infinite-dimensional function. We demonstrate the feasibility of bypassing representation learning and directly employing generation models on infinite-dimensional functions. These findings shed light on novel approaches to 3d object generation and pave the way for more efficient and effective modeling techniques.

嘉宾简介

Biao Zhang is researcher specializing in 3D computer vision. He earned his Bachelor's and Master's degreees from Xi'an Jiaotong University and completed his PhD at KAUST. His recent research focus is on 3D generative models, with publications in CVPR, NeurIPS, ICLR, and SIGGRAPH.

个人主页

主持人简介

施逸飞

国防科技大学

施逸飞，国防科技大学副教授，国防科技大学-普林斯顿大学联合培养博士。研究方向为三维视觉、计算机图形学和智能机器人。入选中国科协青年人才托举工程、中国图学学会青年托举计划、湖湘青年英才、国防科技大学高层次创新人才。发表学术论文40余篇，包括IEEE T-PAMI、ACM TOG、SIGGRAPH、CVPR、ICCV、ECCV (Oral)等顶级期刊会议第一作者论文11篇，CCF A类13篇。主持湖南省优秀青年基金、国家自然科学基金青年项目、国防科技大学基石基金等项目8项，参与国家自然科学基金重点项目、国家重点研发计划等重要项目10余项，担任某重大科研任务主任设计师。获ACM Changsha Chapter新星奖、国防科技大学青年创新奖一等奖等。GAMES执行委员会委员。

GAMES网站: http://games-cn.org