Date of Award
Spring 2023
Document Type
Dissertation
Degree Name
Doctor of Philosophy (PhD)
Department
Electrical and Computer Engineering
First Advisor
Hayat, Majeed
Second Advisor
Medeiros, Henry
Third Advisor
Ababei, Cristinel
Abstract
This dissertation addresses three limitations of deep learning methods in image and video understanding-based machine vision applications. Firstly, although deep convolutional neural networks (CNNs) are efficient for image recognition applications such as object detection and segmentation, they perform poorly under perspective distortions. In real-world applications, the camera perspective is a common problem that we can address by annotating large amounts of data, thus limiting the applicability of the deep learning models. Secondly, the typical approach for single-camera tracking problems is to use separate motion and appearance models, which are expensive in terms of computations and training data requirements. Finally, conventional multi-camera video understanding techniques use supervised learning algorithms to determine temporal relationships among objects. In large-scale applications, these methods are also limited by the requirement of extensive manually annotated data and computational resources.To address these limitations, we develop an uncertainty-aware self-supervised learning (SSL) technique that captures a model's instance or semantic segmentation uncertainty from overhead images and guides the model to learn the impact of the new perspective on object appearance. The test-time data augmentation-based pseudo-label refinement technique continuously trains a model until convergence on new perspective images. The proposed method can be applied for both self-supervision and semi-supervision, thus increasing the effectiveness of a deep pre-trained model in new domains. Extensive experiments demonstrate the effectiveness of the SSL technique in both object detection and semantic segmentation problems. In video understanding applications, we introduce simultaneous segmentation and tracking as an unsupervised spatio-temporal latent feature clustering problem. The jointly learned multi-task features leverage the task-dependent uncertainty to generate discriminative features in multi-object videos. Experiments have shown that the proposed tracker outperforms several state-of-the-art supervised methods. Finally, we proposed an unsupervised multi-camera tracklet association (MCTA) algorithm to track multiple objects in real-time. MCTA leverages the self-supervised detector model for single-camera tracking and solves the multi-camera tracking problem using multiple pair-wise camera associations modeled as a connected graph. The graph optimization method generates a global solution for partially or fully overlapping camera networks.