Dept. of Electrical Engineering & Computer Science
Amanpreet S. Walia is a 4th-year Computer Engineering student at Lassonde School of Engineering, York University. Being motivated to do research in Computer Vision, Amanpreet is spending the summer exploring the design and building of video database for dynamic textures in Dr. Richard’s laboratory. This project involves development and testing of computer vision algorithms for dynamic scene recognition. Specifically, Amanpreet is responsible for formulation of texture categories, design and implementation of tools for collecting videos and annotate them to be utilized by the research community for various tasks. Collected video are annotated for textures to provide accurate ground truths.This project will greatly enhance understanding about construction of databases used in computer vision, where evaluation with respect to significant databases has become a standard methodology.
Building the world's largest Dynamic Scenes Videos Database
Scene classification is a fundamental challenge to the goal of automated visual perception. Although humans are proficient at perceiving and understanding scenes, making computers do the same poses a challenge due to the wide range of variations in scene appearance. Currently, there are a variety of algorithms available to attack this problem; however, algorithmic advances in this area are being held back by the lack of adequate video databases on which to train and test. This project directly addresses this shortcoming by building a video database to support the training and testing of dynamic scene recognition algorithms. The main goal of this project involves developing a large dataset with videos of a variety of dynamic scenes. This task can be categorized into the formulation of scene categories, design and implementation of tools for collection of videos and annotate them to be utilized by the research community. My research focuses on collecting high-quality videos from the web based on formulated scene categories, as well as annotating and analyzing them. To achieve these results, popular video websites are crawled for videos based on keywords selected according to the desired categories to automatically download videos of various dynamic scenes. The collected videos are cleaned through semi-automated processing to gather useful frames and segmented into the components that make up the scene to provide ground truth to segmentation algorithms. These videos are statistically analyzed to highlight various aspects of the database. By the end of summer, this project will be completed providing a new dataset to the computer vision community as a novel benchmark that will help researchers from around the world in contributing to the advancement of computer vision.