Currently, a popular approach to image classification uses the deep Transformer architecture. In a Transformer, the attention mechanism enables the model to learn efficiently with fewer computational resources than the convolutional neural networks (CNNs). In this thesis, we study the sparse attention mechanism widely used in the Transformers developed specifically...
This dissertation addresses the problem of video labeling at both the frame and pixel levels using deep learning. For pixel-level video labeling, we have studied two problems: i) Spatiotemporal video segmentation and ii) Boundary detection and boundary flow estimation. For the problem of spatiotemporal video segmentation, we have developed recurrent...
This dissertation addresses object recognition in challenging settings, where distinct object classes are visually very similar (e.g., species of birds and insects) and/or access to training examples of object classes is limited (e.g., due to the associated high costs of data annotation). In this dissertation, we present a variety of...
This thesis is about visual relationship detection. This is an important task in computer vision. The goal is to detect all visual relationships in a given image between objects. This thesis presents a new approach to this problem. Our approach does not use an object detector as a common pre-processing...
This paper addresses the high model complexity and overconfident frame labeling of state-of-the-art (SOTA) action segmenters. Their complexity is typically justified by the need to sequentially refine action segmentation through multiple stages of a deep architecture. However, this multistage refinement does not take into account uncertainty of frame labeling predicted...
This dissertation addresses few-shot object segmentation in images. The goal of segmentation is to label every image pixel with a class of the object occupying that pixel, where the class may represent a semantic object category or instance. In few-shot segmentation, training and test datasets have different classes. Every new...
Recognizing human actions in videos is a long-standing problem in computer vision with a wide range of applications including video surveillance, content retrieval, and sports analysis. This thesis focuses on addressing efficiency and robustness of video classification in unconstrained real-world settings. The thesis work can be broadly divided into four...
This dissertation addresses the problem of recognizing human activities in videos. Our focus is on activities with stochastic structure, where the activities are characterized by variable space-time arrangements of actions, and conducted by a variable number of actors. These activities occur frequently in sports and surveillance videos. They may appear...
This dissertation addresses the problem of semantic labeling of image pixels. In the course of our work, we considered different types of semantic labels, including object classes (e.g., car, person), 3D depth values (in the range 0 to 80 meters), and affordance classes (e.g., walkable, sittable). Semantic pixel labeling is...
Constructing a panorama from a set of videos is a long-standing problem in computer vision. A panorama represents an enhanced still-image representation of an entire scene captured in a set of videos, where each video shows only a part of the scene. Importantly, a panorama shows only the scene background,...