Theses and Dissertations from UMD
Permanent URI for this communityhttp://hdl.handle.net/1903/2
New submissions to the thesis/dissertation collections are added automatically as they are received from the Graduate School. Currently, the Graduate School deposits all theses and dissertations from a given semester after the official graduation date. This means that there may be up to a 4 month delay in the appearance of a give thesis/dissertation in DRUM
More information is available at Theses and Dissertations at University of Maryland Libraries.
Browse
5 results
Search Results
Item Leveraging Deep Generative Models for Estimation and Recognition(2023) PNVR, Koutilya; Jacobs, David W.; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Generative models are a class of statistical models that estimate the joint probability distribution on a given observed variable and a target variable. In computer vision, generative models are typically used to model the joint probability distribution of a set of real image samples assumed to be on a complex high-dimensional image manifold. The recently proposed deep generative architectures such as Variational Autoencoders (VAEs), Generative Adversarial Networks (GANs), and diffusion models (DMs) were shown to generate photo-realistic images of human faces and other objects. These generative models also became popular for other generative tasks such as image editing, text-to-image, etc. As appealing as the perceptual quality of the generated images has become, the use of generative models for discriminative tasks such as visual recognition or geometry estimation has not been well studied. Moreover, with different kinds of powerful generative models getting popular lately, it's important to study their significance in other areas of computer vision. In this dissertation, we demonstrate the advantages of using generative models for applications that go beyond just photo-realistic image generation: Unsupervised Domain Adaptation (UDA) between synthetic and real datasets for geometry estimation; Text-based image segmentation for recognition. In the first half of the dissertation, we propose a novel generative-based UDA method for combining synthetic and real images when training networks to determine geometric information from a single image. Specifically, we use a GAN model to map both synthetic and real domains into a shared image space by translating just the domain-specific task-related information from respective domains. This is connected to a primary network for end-to-end training. Ideally, this results in images from two domains that present shared information to the primary network. Compared to previous approaches, we demonstrate an improved domain gap reduction and much better generalization between synthetic and real data for geometry estimation tasks such as monocular depth estimation and face normal estimation. In the second half of the dissertation, we showcase the power of a recent class of generative models for improving an important recognition task: text-based image segmentation. Specifically, large-scale pre-training tasks like image classification, captioning, or self-supervised techniques do not incentivize learning the semantic boundaries of objects. However, recent generative foundation models built using text-based latent diffusion techniques may learn semantic boundaries. This is because they must synthesize intricate details about all objects in an image based on a text description. Therefore, we present a technique for segmenting real and AI-generated images using latent diffusion models (LDMs) trained on internet-scale datasets. First, we show that the latent space of LDMs (z-space) is a better input representation compared to other feature representations like RGB images or CLIP encodings for text-based image segmentation. By training the segmentation models on the latent z-space, which creates a compressed representation across several domains like different forms of art, cartoons, illustrations, and photographs, we are also able to bridge the domain gap between real and AI-generated images. We show that the internal features of LDMs contain rich semantic information and present a technique in the form of LD-ZNet to further boost the performance of text-based segmentation. Overall, we show up to 6% improvement over standard baselines for text-to-image segmentation on natural images. For AI-generated imagery, we show close to 20% improvement compared to state-of-the-art techniques.Item Sparse and Deep Representations for Face Recognition and Object Detection(2019) Xu, Hongyu; Chellappa, Rama; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Face recognition and object detection are two very fundamental visual recognition applications in computer vision. How to learn “good” feature representations using machine learning has become the cornerstone of perception-based systems. A good feature representation is often the one that is robust and discriminative to multiple instances of the same category. Starting from features such as intensity, histogram etc. in the image, followed by hand-crafted features, to the most recent sophisticated deep feature representations, we have witnessed the remarkable improvement in the ability of a feature learning algorithm to perform pattern recognition tasks such as face recognition and object detection. One of the conventional feature learning methods, dictionary learning has been proposed to learn discriminative and sparse representations for visual recognition. These dictionary learning methods can learn both representative and discriminative dictionaries, and the associated sparse representations are effective for vision tasks such as face recognition. More recently, deep features have been widely adopted by the computer vision community owing to the powerful deep neural network, which is capable of distilling information from high dimensional input spaces to a low dimensional semantic space. The research problems which comprise this dissertation lie at the cross section of conventional feature and deep feature learning approaches. Thus, in this dissertation, we study both sparse and deep representations for face recognition and object detection. First, we begin by studying the topic of spare representations. We present a simple thresholded feature learning algorithm under sparse support recovery. We show that under certain conditions, the thresholded feature exactly recovers the nonzero support of the sparse code. Secondly, based on the theoretical guarantees, we derive the model and algorithm named Dictionary Learning for Thresholded Features (DLTF), to learn the dictionary that is optimized for the thresholded feature. The DLTF dictionaries are specifically designed for using the thresholded feature at inference, which prioritize simplicity, efficiency, general usability and theoretical guarantees. Both synthetic simulations and real-data experiments (i.e. image clustering and unsupervised hashing) verify the competitive quantitative results and remarkable efficiency of applying thresholded features with DLTF dictionaries. Continuing our focus on investigating the sparse representation and its application to computer vision tasks, we address the sparse representations for unconstrained face verification/recognition problem. In the first part, we address the video-based face recognition problem since it brings more challenges due to the fact that the videos are often acquired under significant variations in poses, expressions, lighting conditions and backgrounds. In order to extract representations that are robust to these variations, we propose a structured dictionary learning framework. Specifically, we employ dictionary learning and low-rank approximation methods to preserve the invariant structure of face images in videos. The learned structured dictionary is both discriminative and reconstructive. We demonstrate the effectiveness of our approach through extensive experiments on three video-based face recognition datasets. Recently, template-based face verification has gained more popularity. Unlike traditional verification tasks, which evaluate on image-to-image or video-to-video pairs, template-based face verification/recognition methods can exploit training and/or gallery data containing a mixture of both images or videos from the person of interest. In the second part, we propose a regularized sparse coding approach for template-based face verification. First, we construct a reference dictionary, which represents the training set. Then we learn the discriminative sparse codes of the templates for verification through the proposed template regularized sparse coding approach. Finally, we measure the similarity between templates. However, in real world scenarios, training and test data are sampled from different distributions. Therefore, we also extend the dictionary learning techniques to tackle the domain adaptation problem, where the data from the training set (source domain) and test set (target domain) have different underlying distributions (domain shift). We propose a domain-adaptive dictionary learning framework to model the domain shift by generating a set of intermediate domains. These intermediate domains bridge the gap between the source and target domains. Specifically, we not only learn a common dictionary to encode the domain-shared features but also learn a set of domain specific dictionaries to model the domain shift. This separation enables us to learn more compact and reconstructive dictionaries for domain adaptation. The domain-adaptive features for recognition are finally derived by aligning all the recovered feature representations of both source and target along the domain path. We evaluate our approach on both cross-domain face recognition and object classification tasks. Finally, we study another fundamental problem in computer vision: generic object detection. Object detection has become one of the most valuable pattern recognition tasks, with great benefits in scene understanding, face recognition, action recognition, robotics and self-driving vehicles, etc. We propose a novel object detector named "Deep Regionlets" by blending deep learning and the traditional regionlet method. The proposed framework "Deep Regionlets" is able to address the limitations of traditional regionlet methods, leading to significant precision improvement by exploiting the power of deep convolutional neural networks. Furthermore, we conduct a detailed analysis of our approach to understand its merits and properties. Extensive experiments on two detection benchmark datasets show that the proposed deep regionlet approach outperforms several state-of-the-art competitors.Item Expressive Knowledge Resources in Probabilistic Models(2014) Hu, Yuening; Boyd-Graber, Jordan; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Understanding large collections of unstructured documents remains a persistent problem. Users need to understand the themes of a corpus and to explore documents of interest. Topic models are a useful and ubiquitous tool to discover the main themes (namely topics) of the corpus. Topic models have been successfully applied in natural language processing, computer vision, information retrieval, cognitive science, etc. However, the discovered topics are not always meaningful: some topics confuse two or more themes into one topic; two different topics can be near duplicates; and some topics make no sense at all. Adding knowledge resources into topic models can improve the topics. However, how to encode knowledge into topic models and where to find these knowledge resources remain two scientific challenges. To address these problems, this thesis presents tree-based topic models to encode prior knowledge, a mechanism incorporating knowledge from untrained users, a polylingual tree-based topic model based on existing dictionaries as knowledge resources, an exploration of regularizing spectral methods to encode prior knowledge into topic models, and a model for automatically building hierarchies of prior knowledge for topic models. To encode knowledge resources into topic models, we first present tree-based topic models, where correlations between word types are modeled as a prior tree and applied to topic models. We also develop more efficient inference algorithms for tree- based topic models. Experiments on multiple corpora show that efficiency is greatly improved on different number of topics, number of correlations and vocabulary size. Because users decide whether the topics are useful or not, users' feedback is necessary for effective topic modeling. We thus propose a mechanism for giving normal users a voice to topic models by encoding users' feedback as correlations between word types into tree-based topic models. This framework, interactive topic modeling (ITM), allows untrained users to encode their feedback easily and iteratively into the topic models. We validate the framework both with simulated and real users and discuss strategies for improving the user experience to adapt models to what users need. Existing knowledge resources such as dictionaries can also improve the model. We propose polylingual tree-based topic models based on bilingual dictionaries and apply this model to domain adaptation for statistical Machine Translation. We derive three different inference schemes and evaluate the efficacy of our model on a Chinese to English translation system, and obtain up to 1.2 BLEU improvement over the machine translation baseline. This thesis further explores an alternative way--regularizing spectral methods for topic models--to encode prior knowledge into topic models. Spectral methods offer scalable alternatives to Markov chain Monte Carlo and expectation maximization. However, these new methods lack the priors that are associated with probabilistic models. We examine Arora et al.'s anchor algorithm for topic models and encode prior knowledge by regularizing the anchor algorithm to improve the interpretability and generalizability of topic models. Because existing knowledge resources are limited and because obtaining the knowledge from users is expensive and time-consuming, automatic techniques should also be considered to extract knowledge from the corpus. This thesis further presents a Bayesian hierarchical clustering technique with the Beta coalescent, which provides a possible way to build up the prior tree automatically. Because of its computational complexity, we develop new sampling schemes using sequential Monte carlo and Dirichlet process mixture models, which render the inference practical and efficient. This thesis explores sources of prior knowledge, presents different ways to encode these expressive knowledge resources into probabilistic topic models, and also applies these models in translation domain adaptation. We also discuss further extensions in a bigger picture of interactive machine learning techniques and domain adaptation for downstream tasks.Item DOMAIN ADAPTIVE OBJECT RECOGNITION AND DETECTION(2013) Mirrashed, Fatemeh; Davis, Larry S; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)Discriminative learning algorithms rely on the assumption that training and test data are drawn from the same marginal probability distribution. In real world applications, however, this assumption is often violated and results in a significant performance drop. We often have sufficient labeled training data from single or multiple "source" domains but wish to learn a classifier which performs well on a "target" domain with a different distribution and no labeled training data. In visual object detection, for example, where the goal is to locate the objects of interest in a given image, it may be infeasible to collect training data to model the enormous variety of possible combinations of pose, background, resolution, and lighting conditions affecting object appearance. Thus, we generally expect to encounter instances or domains at test time for which we have seen little or no training data. To this end, we first propose a framework for domain adaptive object recognition and detection using Transfer Component Analysis, an unsupervised domain adaptation and dimensionality reduction technique. The idea is to obtain a transformation in feature space to a latent subspace that reduces the distance between the source and target data distributions. We evaluate the effectiveness of this approach for vehicle detection using video frames from 50 different surveillance cameras. Next, we explore the problem of extreme class imbalance present when performing fully unsupervised domain adaptation for object detection. The main challenge arises from the fact that images in unconstrained settings are mostly occupied by the background (negative class). Therefore, random sampling will not be effective in obtaining a sufficient number of positive samples from the target domain, which is required by any adaptation method. We propose a variation of co-learning technique that automatically constructs a more balanced set of samples from the target domain. We compare the performance of our technique with other approaches such as unbiased learning from multiple datasets and self-learning. Finally, we propose a novel approach for unsupervised domain adaptation. Our method learns a set of binary attributes for classification that captures the structural information of the data distribution in the target domain itself. The key insight is finding attributes that are discriminative across categories and predictable across domains. We formulate our optimization problem to learn these attributes and the classifier jointly. We evaluate the performance of our method on a wide range of tasks including cross-domain object recognition and sentiment analysis on textual data both in inductive and transductive settings. We achieve a performance that significantly exceeds the state-of-the-art results on standard benchmarks. In many cases we reach the same-domain performance, the upper bound, in unsupervised domain adaptation scenarios.Item NON-LINEAR AND SPARSE REPRESENTATIONS FOR MULTI-MODAL RECOGNITION(2013) Nguyen, Hien Van; Nguyen, Hien V; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)In the first part of this dissertation, we address the problem of representing 2D and 3D shapes. In particular, we introduce a novel implicit shape representation based on Support Vector Machine (SVM) theory. Each shape is represented by an analytic decision function obtained by training an SVM, with a Radial Basis Function (RBF) kernel, so that the interior shape points are given higher values. This empowers support vector shape (SVS) with multifold advantages. First, the representation uses a sparse subset of feature points determined by the support vectors, which significantly improves the discriminative power against noise, fragmentation and other artifacts that often come with the data. Second, the use of the RBF kernel provides scale, rotation, and translation invariant features, and allows a shape to be represented accurately regardless of its complexity. Finally, the decision function can be used to select reliable feature points. These features are described using gradients computed from highly consistent decision functions instead of conventional edges. Our experiments on 2D and 3D shapes demonstrate promising results. The availability of inexpensive 3D sensors like Kinect necessitates the design of new representation for this type of data. We present a 3D feature descriptor that represents local topologies within a set of folded concentric rings by distances from local points to a projection plane. This feature, called as Concentric Ring Signature (CORS), possesses similar computational advantages to point signatures yet provides more accurate matches. CORS produces compact and discriminative descriptors, which makes it more robust to noise and occlusions. It is also well-known to computer vision researchers that there is no universal representation that is optimal for all types of data or tasks. Sparsity has proved to be a good criterion for working with natural images. This motivates us to develop efficient sparse and non-linear learning techniques for automatically extracting useful information from visual data. Specifically, we present dictionary learning methods for sparse and redundant representations in a high-dimensional feature space. Using the kernel method, we describe how the well-known dictionary learning approaches such as the method of optimal directions and KSVD can be made non-linear. We analyse their kernel constructions and demonstrate their effectiveness through several experiments on classification problems. It is shown that non-linear dictionary learning approaches can provide significantly better discrimination compared to their linear counterparts and kernel PCA, especially when the data is corrupted by different types of degradations. Visual descriptors are often high dimensional. This results in high computational complexity for sparse learning algorithms. Motivated by this observation, we introduce a novel framework, called sparse embedding (SE), for simultaneous dimensionality reduction and dictionary learning. We formulate an optimization problem for learning a transformation from the original signal domain to a lower-dimensional one in a way that preserves the sparse structure of data. We propose an efficient optimization algorithm and present its non-linear extension based on the kernel methods. One of the key features of our method is that it is computationally efficient as the learning is done in the lower-dimensional space and it discards the irrelevant part of the signal that derails the dictionary learning process. Various experiments show that our method is able to capture the meaningful structure of data and can perform significantly better than many competitive algorithms on signal recovery and object classification tasks. In many practical applications, we are often confronted with the situation where the data that we use to train our models are different from that presented during the testing. In the final part of this dissertation, we present a novel framework for domain adaptation using a sparse and hierarchical network (DASH-N), which makes use of the old data to improve the performance of a system operating on a new domain. Our network jointly learns a hierarchy of features together with transformations that rectify the mismatch between different domains. The building block of DASH-N is the latent sparse representation. It employs a dimensionality reduction step that can prevent the data dimension from increasing too fast as traversing deeper into the hierarchy. Experimental results show that our method consistently outperforms the current state-of-the-art by a significant margin. Moreover, we found that a multi-layer {DASH-N} has an edge over the single-layer DASH-N.