In the present age of advanced computer vision, the necessity of (user-annotated) data is a key factor in image & video understanding. Recent success of deep learning on large scale data has only acted as a catalyst. There are certain problems that exist in this regard: 1) scarcity of (annotated) data, 2) need of expensive manual annotation, 3) problem of change in domain, 4) knowledge base not exhaustive. To make efficient learning systems, one has to be prepared to deal with such diverse set of problems. In terms of data availability, extensive manual annotation can be beneficial in obtaining category specific knowledge. Even then, learning efficient representation for the related task is challenging and requires special attention. On the other hand, when labelled data is scarce, learning category specific representation itself becomes challenging. In this work, I investigate data driven approaches that cater to traditional supervised learning setup as well as an extreme case of data scarcity where no data from test classes are available during training, known as zero-shot learning. First, I look into supervised learning setup with ample annotations and propose efficient dictionary learning technique for better learning of data representation for the task of action classification in images & videos. Then I propose robust mid-level feature representations for action videos that are equally effective in traditional supervised learning as well as zero-shot learning. Finally, I come up with novel approach that cater to zero-shot learning specifically. Thorough discussions followed by experimental validations establish the worth of these novel techniques in solving computer vision related tasks under varying data-dependent scenarios.
|Titolo della tesi:||Data Driven Approaches for Image & Video Understanding: from Traditional to Zero-shot Supervised Learning|
|Data di discussione:||21-feb-2019|
|Appare nelle tipologie:||Tesi di dottorato|