CVPR 2020

A Comprehensive Tutorial on Video Modeling

Seattle, USA
Sunday Morning, June 14, 2020



This is a tutorial on video modeling organized by Amazon AWS. Our target audience includes students, researchers and engineers, who are interested in learning the recent advances in video modeling, performing research and applying them to real-world problems.

In this tutorial, we will have six technical sessions. We first briefly introduce the problem of human activity understanding in videos, including its input data, common tasks, popular models, and the open challenges. Following it, we dive deep into the technical details, and review recent video modeling methods in a chronological manner. We also introduce GluonCV video model zoo, which has coverage for popular video models and datasets with extensive tutorials. In order to train deep video models efficiently, we introduce an efficient video reader, Decord. Decord provides easy-to-use python interface for video slicing and high efficiency over existing video readers like OpenCV and PyAV. Then, we transit from modeling to deployment, and introduce the best practices we use to deploy video models to production ready devices, such as Jetson Nano/Xavier. In the end, we walk through the diverse set of video research being done at AWS, including tracking, pose estimation, activity classification and action detection.


08:30 - 08:45 : Opening

08:45 - 09:30 : Introduction to Human Activity Understanding in Videos by Yuanjun Xiong [slides] [talk]

09:30 - 10:10 : A Chronological Review of Recent SoTA and Beyond by Yi Zhu [slides] [talk] [GluonCV]

10:10 - 10:30 : Decord: An Efficient Video Reader for Deep Learning by Yi Zhu [slides] [talk] [notebook]

10:30 - 10:45 : Break

10:45 - 11:30 : Deploy Video Models by Zhi Zhang [slides] [talk] [notebook]

11:30 - 12:15 : A Journey Through Video Research at AWS by Joseph Tighe [slides] [talk]

12:15 - 12:45 : Structured Representations for Video Understanding by Chuang Gan [slides] [talk]

For offline Q&A, please post questions to Google Doc

Organizers: Yi Zhu, Zhi Zhang, Yuanjun Xiong and Mu Li

Please contact Yi Zhu if you have question.