Student research opportunities

Multimedia Event Detection from Videos

Project Code: CECS_650

This project is available at the following levels:
Honours, Summer Scholar, Masters, PhD

Keywords:

multimedia, computer vision, machine learning, data mining

Supervisors:

Dr Lexing Xie
Dr Xuming He

Outline:

We are in the age of youtube and influx of many other forms of video clips. Home videos often gets shot in the order of hours but ended up sitting in shoeboxes for years.

A video of a simple daily event, say, "making a cake", can take many different forms (see picture below). So what constitutes "cake-making", how can we tell it apart, and how can we use what we learn to help people make better, and more interesting videos?

Goals of this project

The goal of this project is to push the boundaries of video event modeling and recognition. Steps can be taken in several directions:
+ Understand human perception and understanding of events. This may involve design of user studies of event perception at different video type, length, and complexity.
+ Categorize events based on their perceived semantic complexity, temporal progression, and other factors of perceptual relevance.
+ Design features to represent events -- what kind of features are suitable for describing an event, a class of events, or a range of event classes?
+ Design novel algorithms to separate different event classes.
+ Derive "rules" from interesting videos and apply them for home video editing.

Requirements/Prerequisites

+ Basic understanding of digital images. A course in computer vision or image processing a plus.
+ Linear algebra and probability. Knowledge in statistical machine learning or data mining a plus.
+ Familiar with one or more programming/scripting languages.
+ Can work independently and in a team.

Student Gain

- Hands-on experience and extensive knowledge with image/video processing tools and algorithms.
- Hands-on experience on popular machine learning algorithm and tools.
- Hands-on experience with large-scale standardized evaluations just as the TRECVID benchmark.
- Perform cutting-edge research on video event understanding and recognition, especially feature design and learning.
- Perform cutting-edge research on machine learning algorithms for temporal sequences.