CS239: ML-driven Video Analytics Systems, Winter 2020Instructor: Ravi Netravali
Course OverviewVideo cameras are pervasive. As camera deployments expand, organizations increasingly rely on analyzing video feeds to guide numerous applications including traffic monitoring, surveillance, and amber alert response. Key to the success of such applications has been recent advances in computer vision, particularly neural network (NN)-based techniques for highly accurate object detection and recognition. Though effective at answering high-level queries about video content, these NN-based pipelines are resource intensive in terms of network and server-side compute overheads. This class will explore a wide range of systems and machine learning optimizations to improve the efficiency of modern video analytics pipelines, without violating latency and query accuracy expectations. Grading
Paper SummariesThis course will be entirely based on research papers. Prior to each class, students will be expected to read the listed research paper(s) and write up a brief summary for each. Paper summaries should be short and include the following components:
Paper summaries should be submitted using this form, and are due by 10pm the night before each class. Students may skip paper summaries for up to 2 papers without any penalty. Also, students presenting a paper need not submit a summary. Paper PresentationsFor each paper, one (or more, depending on enrollment) student will be expected to present the paper and lead the discussion for it. Presentations are 30 minutes, should be “conference style”, and describe the domain and relevant background for the paper, the problem statement and challenges, the solution, results, and potential limitations and improvements. Non-presenters are expected to actively participate in the post-presentation discussions. Presenters are expected to come prepared with discussion points and non-presenters should come with ideas or questions for discussions (based on their paper summaries). Active participation will lead to a lively discussion that will benefit everyone. Research ProjectIn addition to paper reading, this course will also include a quarter-long research project. Students will carry out projects individually. The goal with this research project is not necessarily to implement a research idea, but instead is to work entirely on formulating and motivating a potential research idea. In other words, your goal by the end of the quarter should be to convince the class that the idea you are interested in is 1) worth exploring, 2) challenging, and 3) previously unsolved. The scope of acceptable topics is quite large — anything related to any part of an end-to-end video analytics system (live or retrospective). Research directions should be ambitious, at the level of a conference paper (note that the course focuses only on motivation, not implementation, of a conference-level idea). It is encouraged to begin thinking about project topics early on in the quarter by reviewing the reading list/topics, and discussing with the instructor. The deliverables for the project are:
|