GAMES Webinar 2023 – 290期(隐式神经建模前沿) | Weikai Chen(Tencent America),孟晓旭(腾讯),胡文博(字节跳动PICO)
【GAMES Webinar 2023-290期】(几何专题-隐式神经建模前沿)
报告嘉宾:Weikai Chen(Tencent America)
报告时间:2023年8月31号星期四晚上8:00-8:45(北京时间)
报告题目:A touch on Neural 3D Shape Synthesis and Analysis
报告摘要:
3D shape synthesis and analysis are highly coupled tasks. While reliable syntheses may rely on accurate analysis results, high-quality synthesized shapes could in turn provide more samples for meaningful analysis. This talk will cover two works specializing in shape synthesis and analysis, respectively. The first work Get3DHuman, which is accepted to ICCV’23, focuses on synthesizing high-quality 3D textured humans. Our key observation is that the 3D generator can benefit significantly from cross-dimenson priors learned through both 2D human generators and 3D reconstructors. Get3DHuman greatly outperforms the SOTA approaches and can support a wide range of applications including shape interpolation, shape re-texturing, and single-view reconstruction through latent inversion. The second work, which was published in CVPR’23, strives to extract structured edges from unstructured point clouds. We present a novel neural volumetric edge representation, coded NerVE, that can be seamlessly converted to a versatile piece-wise linear (PWL) curve. NerVE paves the way to analyzing complex 3D shapes by offering a unified strategy for learning all types of free-form curves.
讲者简介:
Weikai Chen is a Lead Research Scientist at Tencent America. He got his Ph.D. from the Department of Computer Science, University of Hong Kong. His research lies in the interplay among computer graphics, computer vision, and machine learning. His work SoftRas has been adopted by Pytorch3D as the core algorithm for differentiable rendering. He has published more than 40 papers on top venues in graphics and vision and received more than 2000 citations in Google Scholar. He received the Best Paper Finalist Award in CVPR’19.
讲者主页:http://chenweikai.github.io/
报告嘉宾:孟晓旭(腾讯)
报告时间:2023年8月31号星期四晚上8:45-9:15(北京时间)
报告题目:3D服装难建模?AI来帮你
报告摘要:
本次报告将会介绍CVPR 2023的文章NeAT: Learning Neural Implicit Surfaces with Arbitrary Topologies from Multi-view Images。神经辐射场结合可微分渲染是当前最流行的三维重建方法,然而当前的可微分渲染方法只支持重建闭合曲面(表面紧凑且没有边界的表面,如球体),不支持重建开放曲面(具有开放边界的表面,如服装、纸张、植物叶片)。因此,各类游戏场景中常见的植物叶片、花朵,以及游戏人物身上的服装等等含有开放曲面的三维模型无法直接通过可微分渲染进行重建,只能依赖手工建模。手工建模增加了游戏内容制作的时间成本和人力成本,为游戏开发带来大量额外开销。我们提出了一种创新的可微分渲染管线,支持从多视角图片重建的任意隐式曲面,并支持快速导出高质量三维模型。
讲者简介:
孟晓旭,腾讯互动娱乐DCC算法研究中心研究员。2015年本科毕业于上海交通大学,2020年于马里兰大学帕克分校取得计算机博士学位。主要研究方向为可微分渲染在三维重建中的应用、光线追踪渲染的降噪及注视点渲染。
讲者主页:https://xmeng525.github.io/xiaoxumeng.github.io/
报告嘉宾:胡文博(字节跳动PICO)
报告时间:2023年8月31号星期四晚上9:15-9:45(北京时间)
报告题目:Tri-MipRF: Tri-Mip Representation for Efficient Anti-Aliasing Neural Radiance Fields
报告摘要:
本次报告的工作被收录于ICCV’23并进行Oral presentation。即使NeRF领域已经取得了长足的发展,但我们仍然面临着一个质量和效率不可兼得的两难境地,比如MipNeRF有着高保真的anti-aliased渲染结果但是需要数天来重建,而Instant-ngp可以在几分钟内完成重建但是当在不同距离合分辨率渲染时会有aliasing和模糊的问题。在这个工作中,我们提出一个新颖的Tri-Mip representation来同时达到高效重建和高质量anti-aliased渲染。其中的关键点是使用三个正交的mipmap来建模pre-filtered 三维特征空间,这样我们就可以借助pre-filtered 2D 特征图进行高效的3D 区域采样来极大地提升渲染质量而不牺牲重建速度。充分的实验表明我们的方法达到了SOTA的渲染质量和重建速度。
讲者简介:
胡文博,字节跳动PICO MR算法工程师。2018年本科毕业于大连理工大学,2022年于香港中文大学获得计算机科学与工程博士学位,研究方向为计算机图形学和3D视觉,近期的研究主要集中在Neural 3D Reconstruction and Rendering。
主持人简介:
杨洁,中国科学院计算技术研究所,任职助理研究员,研究方向为几何处理和几何学习。博士毕业于中科院计算所。研究成果发表在ACM SIGGRAPH/TOG、IEEE TPAMI、NeurIPS等期刊和会议上。担任CVPR, ECCV, ICCV, AAAI, NeurIPS, CGF, IJCV, TVCJ等期刊会议的审稿人。个人主页:http://people.geometrylearning.com/~jieyang/
GAMES主页的“使用教程”中有 “如何观看GAMES Webinar直播?”及“如何加入GAMES微信群?”的信息;
GAMES主页的“资源分享”有往届的直播讲座的视频及PPT等。
观看直播的链接:http://webinar.games-cn.org