大家好 真是好久不见!
随着夏天悄悄地靠近
我们新学期的工作也在火热进行中~
本学期第一次讲座来袭啦
让我们一起看看本次讲座的内容吧
When Depth Estimation Meets
Deep Learning
时间:4月20日 16:20
地点:南实验楼E403
主讲人介绍
孙文秀:商汤科技研究院 研究副总监
本科毕业于南京大学电子科学与工程系,博士毕业于香港科技大学电子与计算机工程系。本科期间曾参加intel嵌入式邀请赛,获得国家一等奖。博士期间去日本国际情报学研究所(NII)进行短期研究访问。研究兴趣包括计算机视觉和像素级图像视频处理,发表计算机视觉顶级会议及期刊30多篇。
2015年加入商汤,现任商汤科技研究副总监,负责计算摄影算法解决方案、深度与运动感知的新技术研究,主要产品输出在手机行业。
讲座内容简介
Depth data is indispensable for reconstructing or understanding 3D scenes. It serves as a key ingredient for applications such as synthetic defocus, autonomous driving, and augmented reality.
Although active 3D sensors (e.g., Lidar, ToF, and structured-light 3D scanner) can be employed, retrieving depth from monocular/stereo cameras is typically a more cost-effective approach.
However, estimating depth from images is inherently under-determined, to regularize the problem, one typically needs handcrafted models characterizing the properties of depth data or scene geometry.
As the recent advances in deep learning, depth estimation is cast as a learning task, leading to state-of-the-art performance. In this talk, I will present our new progress on depth estimation with convolutional neural networks (CNN).
Particularly, I will first introduce cascade residual learning (CRL), our two-stage deep architecture on stereo matching producing high-quality disparity estimates. Observations with CRL inspires us to propose a domain-adaptation approach---zoom and learn (ZOLE)---for training a deep stereo matching algorithm without the ground-truth data of a target domain.
By combining a view synthesis network and the first stage of CRL, we propose single view stereo matching (SVS) for single image depth estimation, with a performance superior to the classic stereo block matching method taking two images as inputs.
Finally, I will present our endeavours when applying our core techniques to the depth-of-field effects on dual-lens smart phones.
扫一扫下图二维码 / 点击阅读原文
即可报名参加讲座