Object-Centered Visual Recognition

Опубликовано 17 августа 2016, 1:02

Machine vision approaches for accurate and efficient visual recognition have the potential to play a transformative role in numerous applications, and while seemingly simple visual tasks continue to elude artificial systems, we are making rapid progress. In this talk I will discuss my research on ΓÇ£object-centeredΓÇ¥ visual recognition, where the goal is to find objects in images and videos and characterize their attributes and activity in detail. Specifically, I will discuss my work on four fundamental aspects of object-centered visual recognition: object detection, pose estimation, object tracking and behavior recognition. First, I will describe our state-of-the-art pedestrian detection approach with a focus on a recent insight that has allowed us to perform accurate multiscale detection in near real time. This approach yields a speedup of 10-100 times over competing methods with only a minor loss in detection accuracy and the underlying theory should be readily applicable to numerous domains. I will also briefly discuss our large-scale benchmarking of pedestrian detection, highlighting current successes and open challenges for the research community. In the second part of the talk I will describe our method for efficient pose estimation, a general tracking by detection system which leverages our research in object detection and pose estimation, and our widely adopted framework for behavior recognition. The approaches I will present are both effective (i.e., accurate and robust) and practical (i.e., computationally efficient and broadly applicable) and these are elements I will highlight throughout the talk.