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

Search Results

Now showing 1 - 8 of 8

Activity Detection in Untrimmed Videos
(2023) Gleason, Joshua D; Chellappa, Rama; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
In this dissertation, we present solutions to the problem of activity detection in untrimmed videos, where we are interested in identifying both when and where various activity instances occur within an unconstrained video. Advances in machine learning, particularly the widespread adoption of deep learning-based methods have yielded robust solutions to a number of historically difficult computer vision application domains. For example, recent systems for object recognition and detection, facial identification, and a number of language processing applications have found widespread commercial success. In some cases, such systems have been able to outperform humans. The same cannot be said for the problem of activity detection in untrimmed videos. This dissertation describes our investigation and innovative solutions for the challenging problem of real-time activity detection in untrimmed videos. The main contributions of our work are the introduction of multiple novel activity detection systems that make strides toward the goal of commercially viable activity detection. The first work introduces a proposal mechanism based on divisive hierarchical clustering of objects to produce cuboid activity proposals, followed by a classification and temporal refinement step. The second work proposes a chunk-based processing mechanism and explores the tradeoff between tube and cuboid proposals. The third work explores the topic of real-time activity detection and introduces strategies for achieving this performance. The final work provides a detailed look into multiple novel extensions that improve upon the state-of-the-art in the field.
LONG-TERM TEMPORAL MODELING FOR VIDEO ACTION UNDERSTANDING
(2021) Yang, Xitong; Davis, Larry S; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
The tremendous growth in video data, both on the internet and in real life, has encouraged the development of intelligent systems that can automatically analyze video contents and understand human actions. Therefore, video understanding has been one of the fundamental research topics in computer vision.Encouraged by the success of deep neural networks on image classification, many efforts have been made in recent years to extend the deep networks to video understanding. However, new challenges arise when the temporal characteristic of videos is taken into account. In this dissertation, we study two long-standing problems that play important roles in effective temporal modeling in videos: (1) How to extract motion information from raw video frames? (2) How to capture long-range dependencies in time and model their temporal dynamics? To address the above issues, we first introduce hierarchical contrastive motion learning, a novel self-supervised learning framework to extract effective motion representations from raw video frames. Our approach progressively learns a hierarchy of motion features, from low-level pixel movements to higher-level semantic dynamics, in a fully self-supervised manner.Next, we investigate the self-attention mechanism for long-range temporal modeling, and demonstrate that the entangled modeling of spatio-temporal information fails to capture temporal relationships among frames explicitly. To this end, we propose Global Temporal Attention (GTA), which performs global temporal attention on top of spatial attention in a decoupled manner. Unlike conventional self-attention that computes an instance-specific attention matrix, GTA directly learns a global attention matrix that is intended to encode temporal structures that generalize across different samples. While the performance of video action recognition has been significantly improved by the aforementioned methods, they are still restricted to model temporal information within short clips. To overcome this limitation, we introduce a collaborative memory mechanism that encodes information across multiple sampled clips of a video at each training iteration. Our proposed framework is end-to-end trainable and significantly improves the accuracy of video classification at a negligible computational overhead. Finally, we present a spatio-temporal progressive learning framework (STEP) for spatio-temporal action detection. Our approach performs a multi-step optimization process that progressively refines the initial proposals towards the final solution. In this way, our approach can effectively make use of long-term temporal information by handling the spatial displacement problem in long action tubes.
Improving Efficiency and Generalization of Visual Recognition
(2018) Yu, Ruichi; Davis, Larry S; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
Deep Neural Networks (DNNs) are heavy in terms of their number of parameters and computational cost. This leads to two major challenges: first, training and deployment of deep networks are expensive; second, without tremendous annotated training data, which are very costly to obtain, DNNs easily suffer over-fitting and have poor generalization. We propose approaches to these two challenges in the context of specific computer vision problems to improve their efficiency and generalization. First, we study network pruning using neuron importance score propagation. To reduce the significant redundancy in DNNs, we formulate network pruning as a binary integer optimization problem which minimizes the reconstruction errors on the final responses produced by the network, and derive a closed-form solution to it for pruning neurons in earlier layers. Based on our theoretical analysis, we propose the Neuron Importance Score Propagation (NISP) algorithm to propagate the importance scores of final responses to every neuron in the network, then prune neurons in the entire networks jointly. Second, we study visual relationship detection (VRD) with linguistic knowledge distillation. Since the semantic space of visual relationships is huge and training data is limited, especially for long-tail relationships that have few instances, detecting visual relationships from images is a challenging problem. To improve the predictive capability, especially generalization on unseen relationships, we utilize knowledge of linguistic statistics obtained from both training annotations (internal knowledge) and publicly available text, e.g., Wikipedia (external knowledge) to regularize visual model learning. Third, we study the role of context selection in object detection. We investigate the reasons why context in object detection has limited utility by isolating and evaluating the predictive power of different context cues under ideal conditions in which context provided by an oracle. Based on this study, we propose a region-based context re-scoring method with dynamic context selection to remove noise and emphasize informative context. Fourth, we study the efficient relevant motion event detection for large-scale home surveillance videos. To detect motion events of objects-of-interest from large scale home surveillance videos, traditional methods based on object detection and tracking are extremely slow and require expensive GPU devices. To dramatically speedup relevant motion event detection and improve its performance, we propose a novel network for relevant motion event detection, ReMotENet, which is a unified, end-to-end data-driven method using spatial-temporal attention-based 3D ConvNets to jointly model the appearance and motion of objects-of-interest in a video. In the last part, we address the recognition of agent-in-place actions, which are associated with agents who perform them and places where they occur, in the context of outdoor home surveillance. We introduce a representation of the geometry and topology of scene layouts so that a network can generalize from the layouts observed in the training set to unseen layouts in the test set. This Layout-Induced Video Representation (LIVR) abstracts away low-level appearance variance and encodes geometric and topological relationships of places in a specific scene layout. LIVR partitions the semantic features of a video clip into different places to force the network to learn place-based feature descriptions; to predict the confidence of each action, LIVR aggregates features from the place associated with an action and its adjacent places on the scene layout. We introduce the Agent-in-Place Action dataset to show that our method allows neural network models to generalize significantly better to unseen scenes.
Video Understanding with Deep Networks
(2018) Ng, Yue Hei; Davis, Larry S; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
Video understanding is one of the fundamental problems in computer vision. Videos provide more information to the image recognition task by adding a temporal component through which motion and other information can be additionally used. Encouraged by the success of deep convolutional neural networks (CNNs) on image classification, we extend the deep convolutional networks to video understanding by modeling both spatial and temporal information. To effectively utilize deep networks, we need a comprehensive understanding of convolutional neural networks. We first study the network on the domain of image retrieval. We show that for instance-level image retrieval, lower layers often perform better than the last layers in convolutional neural networks. We present an approach for extracting convolutional features from different layers of the networks and adopt VLAD encoding to encode features into a single vector for each image. Our work provides guidance for transferring deep convolutional networks to other tasks. We then propose and evaluate several deep neural network architectures to combine image information across a video over longer time periods than previously attempted. We propose two methods capable of handling full length videos. The first method explores various convolutional temporal feature pooling architectures, examining the various design choices which need to be made when adapting a CNN for this task. The second proposed method explicitly models the video as an ordered sequence of frames. For this purpose, we employ a recurrent neural network that uses Long Short-Term Memory (LSTM) cells which are connected to the output of the underlying CNN. Next, we propose a multitask learning model ActionFlowNet to train a single stream network directly from raw pixels to jointly estimate optical flow while recognizing actions with convolutional neural networks, capturing both appearance and motion in a single model. Experiments show that our model effectively learns video representation from motion information on unlabeled videos. While recent deep models for videos show improvement by incorporating optical flow or aggregating high-level appearance across frames, they focus on modeling either the long-term temporal relations or short-term motion. We propose Temporal Difference Networks (TDN) that model both long-term relations and short-term motion from videos. We leverage a simple but effective motion representation: difference of CNN features in our network and jointly modeling the motion at multiple scales in a single CNN.
Multimodal Approaches to Computer Vision Problems
(2017) Reale, Chris; Chellappa, Rama; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
The goal of computer vision research is to automatically extract high-level information from images and videos. The vast majority of this research focuses specifically on visible light imagery. In this dissertation, we present approaches to computer vision problems that incorporate data obtained from alternative modalities including thermal infrared imagery, near-infrared imagery, and text. We consider approaches where other modalities are used in place of visible imagery as well as approaches that use other modalities to improve the performance of traditional computer vision algorithms. The bulk of this dissertation focuses on Heterogeneous Face Recognition (HFR). HFR is a variant of face recognition where the probe and gallery face images are obtained with different sensing modalities. We also present a method to incorporate text information into human activity recognition algorithms. We first present a kernel task-driven coupled dictionary model to represent the data across multiple domains for thermal infrared HFR. We extend a linear coupled dictionary model to use the kernel method to process the signals in a high dimensional space; this effectively enables the dictionaries to represent the data non-linearly in the original feature space. We further improve the model by making the dictionaries task-driven. This allows us to tune the dictionaries to perform well on the classification task at hand rather than the standard reconstruction task. We show that our algorithms outperform algorithms based on standard coupled dictionaries on three datasets for thermal infrared to visible face recognition. Next, we present a deep learning-based approach to near-infrared (NIR) HFR. Most approaches to HFR involve modeling the relationship between corresponding images from the visible and sensing domains. Due to data constraints, this is typically done at the patch level and/or with shallow models to prevent overfitting. In this approach, rather than modeling local patches or using a simple model, we use a complex, deep model to learn the relationship between the entirety of cross-modal face images. We describe a deep convolutional neural network-based method that leverages a large visible image face dataset to prevent overfitting. We present experimental results on two benchmark data sets showing its effectiveness. Third, we present a model order selection algorithm for deep neural networks. In recent years, deep learning has emerged as a dominant methodology in machine learning. While it has been shown to produce state-of-the-art results for a variety of applications, one aspect of deep networks that has not been extensively researched is how to determine the optimal network structure. This problem is generally solved by ad hoc methods. In this work we address a sub-problem of this task: determining the breadth (number of nodes) of each layer. We show how to use group-sparsity-inducing regularization to automatically select these hyper-parameters. We demonstrate the proposed method by using it to reduce the size of networks while maintaining performance for our NIR HFR deep-learning algorithm. Additionally, we demonstrate the generality of our algorithm by applying it to image classification tasks. Finally, we present a method to improve activity recognition algorithms through the use of multitask learning and information extracted from a large text corpora. Current state-of-the-art deep learning approaches are limited by the size and scope of the data set they use to train the networks. We present a multitask learning approach to expand the training data set. Specifically, we train the neural networks to recognize objects in addition to activities. This allows us to expand our training set with large, publicly available object recognition data sets and thus use deeper, state-of-the-art network architectures. Additionally, when learning about the target activities, the algorithms are limited to the information contained in the training set. It is virtually impossible to capture all variations of the target activities in a training set. In this work, we extract information about the target activities from a large text corpora. We incorporate this information into the training algorithm by using it to select relevant object recognition classes for the multitask learning approach. We present experimental results on a benchmark activity recognition data set showing the effectiveness of our approach.
MANIPULATION ACTION UNDERSTANDING FOR OBSERVATION AND EXECUTION
(2015) Yang, Yezhou; Aloimonos, Yiannis; Fermuller, Cornelia; Computer Science; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
Modern intelligent agents will need to learn the actions that humans perform. They will need to recognize these actions when they see them and they will need to perform these actions themselves. We want to propose a cognitive system that interprets human manipulation actions from perceptual information (image and depth data) and consists of perceptual modules and reasoning modules that are in interaction with each other. The contributions of this work are given along two core problems at the heart of action understanding: a.) the grounding of relevant information about actions in perception (the perception - action integration problem), and b.) the organization of perceptual and high-level symbolic information for interpreting the actions (the sequencing problem). At the high level, actions are represented with the Manipulation Action Context-free Grammar (MACFG) , a syntactic grammar and associated parsing algorithms, which organizes actions as a sequence of sub-events. Each sub-event is described by the hand (as well as grasp type), movements (actions) and the objects and tools involved, and the relevant information about these quantities is obtained from biological-inspired perception modules. These modules track the hands and objects and recognize the hand grasp, actions, segmentation, and action consequences. Furthermore, a probabilistic semantic parsing framework based on CCG (Combinatory Categorial Grammar) theory is adopted to model the semantic meaning of human manipulation actions. Additionally, the lesson from the findings on mirror neurons is that the two processes of interpreting visually observed action and generating actions, should share the same underlying cognitive process. Recent studies have shown that grammatical structures underlie the representation of manipulation actions, which are used both to understand and to execute these actions. Analogically, understanding manipulation actions is like understanding language, while executing them is like generating language. Experiments on two tasks, 1) a robot observing people performing manipulation actions, and 2) a robot then executing manipulation actions accordingly, are presented to validate the formalism. The technical parts of this thesis are devoted to the experimental setting of task (1), while the task (2) is given as a live demonstration.
Scene and Video Understanding
(2014) Jain, Arpit; Davis, Larry S; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
There have been significant improvements in the accuracy of scene understanding due to a shift from recognizing objects ``in isolation'' to context based recognition systems. Such systems improve recognition rates by augmenting appearance based models of individual objects with contextual information based on pairwise relationships between objects. These pairwise relations incorporate common sense world knowledge such as co-occurrences and spatial arrangements of objects, temporal consistency, scene layout, etc. However, these relations, even though consistent in the 3D world, change due to viewpoint of the scene. In this thesis, we investigate incorporating contextual information from three different perspectives for scene and video understanding (a) ``what'' contextual relations are useful and ``how'' they should be incorporated into Markov network during inference, (b) jointly solving the segmentation and recognition problem using a multiple segmentation framework based on contextual information in conjunction with appearance matching, and (c) proposing a discriminative spatio-temporal patch based representation for videos which incorporates contextual information for video understanding. Our work departs from traditional view of incorporating context into scene understanding where a fixed model for context is learned. We argue that context is scene dependent and propose a data-driven approach to predict the importance of relationships and construct a Markov network for image analysis based on statistical models of global and local image features. Since all contextual information is not equally important, we also address the related problem of predicting the feature weights associated with each edge of a Markov network for evaluation of context. We then address the problem of fixed segmentation while modeling context by using a multiple segmentation framework and formulating the problem as ``a jigsaw puzzle''. We formulate the labeling problem as segment selection from a pool of segments (jigsaws), assigning each selected segment a class label. Previous multiple segmentation approaches used local appearance matching to select segments in a greedy manner. In contrast, our approach is based on a cost function that combines contextual information with appearance matching. A relaxed form of the cost function is minimized using an efficient quadratic programming solver. Lastly, we propose a new representation for videos based on mid-level discriminative spatio-temporal patches. These patches might correspond to a primitive human action, a semantic object, or perhaps a random but informative spatiotemporal patch in the video. What define these spatiotemporal patches are their discriminative and representative properties. We automatically mine these patches from hundreds of training videos and experimentally demonstrate that these patches establish correspondence across videos. We propose a cost function that incorporates co-occurrence statistics and temporal context along with appearance matching to select subset of these patches for label transfer. Furthermore, these patches can be used as a discriminative vocabulary for action classification.
Modeling Shape, Appearance and Motion for Human Movement Analysis
(2009) Lin, Zhe; Davis, Larry S; Electrical Engineering; Digital Repository at the University of Maryland; University of Maryland (College Park, Md.)
Shape, Appearance and Motion are the most important cues for analyzing human movements in visual surveillance. Representation of these visual cues should be rich, invariant and discriminative. We present several approaches to model and integrate them for human detection and segmentation, person identification, and action recognition. First, we describe a hierarchical part-template matching approach to simultaneous human detection and segmentation combining local part-based and global shape-based schemes. For learning generic human detectors, a pose-adaptive representation is developed based on a hierarchical tree matching scheme and combined with an support vector machine classifier to perform human/non-human classification. We also formulate multiple occluded human detection using a Bayesian framework and optimize it through an iterative process. We evaluated the approach on several public pedestrian datasets. Second, given regions of interest provided by human detectors, we introduce an approach to iteratively estimates segmentation via a generalized Expectation-Maximization algorithm. The approach incorporates local Markov random field constraints and global pose inferences to propagate beliefs over image space iteratively to determine a coherent segmentation. Additionally, a layered occlusion model and a probabilistic occlusion reasoning scheme are introduced to handle inter-occlusion. The approach is tested on a wide variety of real-life images. Third, we describe an approach to appearance-based person recognition. In learning, we perform discriminative analysis through pairwise coupling of training samples, and estimate a set of normalized invariant profiles by marginalizing likelihood ratio functions which reflect local appearance differences. In recognition, we calculate discriminative information-based distances by a soft voting approach, and combine them with appearance-based distances for nearest neighbor classification. We evaluated the approach on videos of 61 individuals under significant illumination and viewpoint changes. Fourth, we describe a prototype-based approach to action recognition. During training, a set of action prototypes are learned in a joint shape and motion space via $k$-means clustering; During testing, humans are tracked while a frame-to-prototype correspondence is established by nearest neighbor search, and then actions are recognized using dynamic prototype sequence matching. Similarity matrices used for sequence matching are efficiently obtained by look-up table indexing. We experimented the approach on several action datasets.

Theses and Dissertations from UMD

Browse

Filters

Settings

Sort By

Results per page

Search Results