主题:Neural Network Compression & Acceleration for Edge Devices

Deep Neural Networks have demonstrated great success in various computer vision tasks. However, their high demands in computing power and memory footprint prohibit most state-of-the-art networks to be employed in edge devices such as smartphones and wearable facilities. This talk is devoted to giving a brief introduction to the prevalent techniques used to compress and accelerate modern neural networks. Moreover, a series of our recent research will be elaborated including network pruning using high-rank feature map, tensor decomposition with knowledge transfer, automatic structure search for a slimmable neural network, network quantization using trainable scale module, etc. The discussions cover our recent publication on IEEE T-PAMI, CVPR, ECCV and IJCAI. The source codes are online available.
  • 纪荣嵘教授,厦门大学教授,国家优秀青年科学基金获得者。主要研究方向为计算机视觉、多媒体技术和机器学习。近年来发表PAMI、IJCV、ACM汇刊、IEEE汇刊、CVPR、NeurIPS等会议长⽂过百篇。论文谷歌学术引用九千余次。曾获2007年微软学者奖、2011年ACM Multimedia最佳论文奖、2015年省自然科学二等奖、2016年教育部技术发明一等奖、2018年省科技进步一等奖、2019年福建省青年科技奖。曾/现主持国家重点研发计划课题,国家自然科学基金。任中国计算机学会A类国际会议IEEE CVPR 2021、ACM Multimedia 2020/ 2019/2018领域主席等。任中国图象图形学学会青工委副主任、中国自动化学会粒计算与多尺度分析专委会副主任、中国计算机学会计算机视觉专委会常务委员、中国计算机学会学术工委委员、教育部电子信息类教指委人工智能专业建设咨询委员会委员。

主题:Single/Multi-modal Object Detection

Object detection is a fundamental research topic in computer vision and has made remarkable progress in recent years. It is widely used in many fields, such as automatic driving, face detection, human-computer interaction and so on. However, natural images usually contain objects of various categories, sizes and semantic confusion, which makes it challenging to accurately locate these objects. This talk will discuss recent research directions and hot topics for the object detection. Furthermore, some related works with deep learning of our research group will be introduced including the single-modal object detection and multi-modal object detection methods. Finally, the challenges and opportunities of object detection in the future will be summarized. This talk includes our recent works at IEEE CVPR, ECCV, and ACM MM.
  • 李宏亮教授,电子科技大学教授,博士生导师,国家杰出青年科学基金获得者,教育部新世纪优秀人才计划入选者,四川省学术和技术带头人。主要研究领域包括图像理解与分析,对象检测与分割,视觉感知模型等。已发表学术论文110余篇,其中包括IEEE Transactions论文50余篇。主持了多项国家和省部级科研项目。目前担任IEEE Transactions on Circuits and Systems for Video Technology编委,Elsevier国际期刊Journal on Visual Communications and Image Representation,Signal Processing: Image Communication编委。2014年IEEE多媒体会议与博览会(ICME2014)的本地组委会主席,2016年IEEE视觉通信和图像处理国际会议(VCIP 2016)技术委员会主席。2017年IEEE智能信号处理与通信系统国际会议大会主席,以及2017环太多媒体会议程序委员会主席。IEEE电路与系统协会VSPC专委会委员。获2018年JVCI杰出服务奖,2019年IEEE-TCSVT最佳编委奖。

主题:Learning Based Optimization

Optimization is a core computational technique for many areas, including machine learning, communication, signal processing, etc. However, in the traditional ways the complexity of solving particular types of problems has lower bounds, which are unbreakable if we confine ourselves to traditional ways. In recent years, there emerges a new optimization strategy, called learning based optimization, which can achieve much faster convergence rates for particular types of data. In this talk, I will introduce the basic ideas, related work, and our recent work.
  • 林宙辰教授,北京大学教授,IAPR/IEEE Fellow,国家杰青,中国图象图形学学会机器视觉专委会主任,中国自动化学会模式识别与机器智能专委会副主任。研究领域为机器学习、计算机视觉、图像处理、模式识别和数值优化。发表论文200余篇,英文专著2本。任CVPR 2014/2016/2019/2020/2021、ICCV 2015、NIPS/NeurIPS 2015/2018/2019/2020、ICML 2020、IJCAI 2020/2021、AAAI 2019/2020和ICLR 2021领域主席,ICPR 2020 Track Chair,IJCV编委。曾任IEEE T. PAMI编委。

主题:Long Range 3D Sensing for High-Speed Autonomous Driving

In this talk, I will present some research work on long range 3D sensing using monocular cameras. High-speed autonomous driving demands 3D perception beyond 200m, which is the typical maximum range for either Lidar or Radar. I will present a new method that combines traditional 3D vision with scene-parsing method to enable accurate vehicle range sensing at over 1000m. A ultra-long range 3D dataset will also be introduced.
  • 杨睿刚教授,2003年于美国北卡罗莱纳大学教堂山分校获博士学位,主修计算机科学。曾获得美国国家基金委员会CAREER奖,美国肯塔基大学计算机系终身教授。 曾任百度研究院机器人和自动驾驶实验室主任。现任嬴彻科技CTO, 杨睿刚博士在包括IJCV、IEEE T-PAMI、SIGGRAPH、CVPR、ICCV在内的计算机视觉和图形学领域顶级期刊和会议上发表论文130 余篇,Google Scholar引用超过万次,H 指数50。

主题:Building a 3D Human from Head to Toe

This talk focuses on novel deep learning methods for constructing ultra-realistic 3D human models. The challenges are multi-fold and metaphorically, from head to toe. For example, even with advanced 3D scanners based on direct or indirect time-of-flight sensors, it remains challenging to recover fine details of facial geometry, let alone hair or shoes. We demonstrate that deep learning, or more precisely, neural modeling and rendering provide a new and viable path: high quality images are much easier to synthesize than geometry. Therefore, we resort to first producing photorealistic images and then use the images to recover and refine human geometry recovery. This talk covers our recent works at IEEE PAMI, ICCV, CVPR and ACM MM where all source codes have been made publicly available.
  • 虞晶怡教授于2000年获美国加州理工学院(Caltech)双学士学位,2005年获美国麻省理工学院(MIT)博士学位。现任上海科技大学信息科学与技术学院教授、执行院长,是上海人工智能咨询委员会委员、叠境数字联合创始人。虞教授长期从事计算机视觉、计算成像、计算机图形学、生物信息学等领域的研究工作,已发表140多篇学术论文, 其中超90篇发表于国际会议CVPR/ICCV/ECCV和期刊TPAMI,已获得20多项PCT发明专利授权。虞教授是美国国家科学基金杰出青年奖(NSF CAREER Award)获得者。他曾组织多个计算机视觉大会,担任IEEE TPAMI、IEEE TIP等多个顶级期刊编委。虞教授担任ICCP 2016、ICPR 2020、WACV 2021以及人工智能顶会IEEE CVPR 2021和ICCV 2025的程序主席。


  • 张长水教授,男,1986年7月毕业于北京大学数学系,获得学士学位。1992年7月毕业于清华大学自动化系,获得博士学位。1992年7月至今在清华大学自动化系工作。现任清华大学自动化系教授,主要研究兴趣包括:机器学习、模式识别、计算视觉等方面。目前是IEEE Fellow, 计算机学会高级会员;担任学术期刊:IEEE Trans. on PAMI 等杂志编委;在国际期刊发表论文130多篇,在顶级会议上发表论文50多篇。