doc. Georgios Tolias, Ph.D.

Recent advances in deep learning have led to astounding performance in visual recognition of generic object and action categories in images and videos, respectively. On the other hand, recognition of categories that are defined with larger granularity, such as in the case of fine-grained or instance-level categorization, is arguably an unsolved problem. The large set of classes, the small inter-class and intra-class variability, along with an exhibited long-tail distribution makes it a very challenging problem. The thesis will depart from conventional deep learning architectures for recognition and will explore ways to learn the visual representation at the desired granularity together with the appropriate matching process. Mid-level representation is identified as one of the key ingredients and it will be explored on two axes. Firstly, in terms of locality, ranging from local to global, and, secondly, in terms of semantics, ranging from low-level features to high-level semantics. The matching process is responsible for inferring visual similarities from the visual representation, and will play a key role in recognition with few training examples. The main focus of the thesis will be on, but will not be limited to, video recognition and understanding, while matching, event detection and alignment are other tasks that the thesis will possibly consider too.