Analytic Training Approach for Object/Action Recognition
The computer vision community is faced with the challenge of devising novel, robust and efficient algorithms to learn models which are helpful in categorizing huge amount of visual data. Recognition algorithms play a pivotal role in commercially available frameworks for automated analysis and detection of objects of interest (OI). Traditionally, supervised learning frameworks have inherent limitations of longer durations for training and finding local maxima that may lead to poor classification accuracy. Recent action recognition schemes have ignored complexity associated with redundant training samples and learning strategies. To cope with inherent limitations of gradient descent approach, model learning can be analytically performed at an extremely fast speed without iterative adjustments. This presentation is mainly concentrated on recent trends in action and activity recognition. We believe that computation, and selection of meaningful features and their use in model learning are equally important for recognition purposes. In this presentation, first we will discuss reasons for analytic training approach and then present a general framework to efficiently identify OI in still images and later extend its application to human action recognition in videos as well. Such scheme can also be implemented in a situation where training data is coming in a serial mode and training needs to be performed in an incremental fashion.
时间:7月5日 10:30-11:30 地点:教十八 223室