of a dissertation at the University of Miami. Dissertation supervised by Professor Mei-Ling Shyu. No. of pages in text. (153) The development in information science has enabled an explosive growth of data, which attracts more and more researchers to engage in the field of big data analytics. Noticeably, in many real-world applications, large amounts of data are imbalanced data since the events of interests occur infrequently. Classification of imbalanced data is an important research problem as lots of real-world datasets have skewed class distributions in which the majority of instances (examples) belong to one class and far fewer instances belong to the others. A classifier induced from an imbalanced dataset is more likely to be biased towards the majority classes and shows very poor classification accuracy on the minority classes. While in many applications, the minority instances actually represent the concept of interest (e.g., fraud in banking operations, abnormal cell in medical data, etc.), and the detection of these rare events has become more important. Despite extensive research efforts, rare event mining remains one of the most challenging problems in information retrieval, especially for multimedia big data. To tackle this challenge, in this dissertation, we propose an extended deep learning approach to achieve promising performance in classifying largely skewed multimedia dataset. Specifically, we investigate the integration of bootstrapping methods and a state-of-the-art deep learning approach, Convolutional Neural Networks (CNNs), with extensive empirical studies. Considering the fact that deep learning approaches such as CNNs are usually computationally expensive, we propose to feed low-level features to CNNs and prove its feasibility in achieving promising performance while saving a lot of training time. Furthermore, since big training datasets are required to train CNNs, we propose to extract features from pre-trained CNN models and feed those features to another full connected neural network. Implementations in big data environments show promising performance of our model in handling big datasets with respect to feasibility and scalability. In order to further improve the classification results and bridge the semantic gap between high-level concepts and low-level visual features, correlation discovery in semantic concept mining is worth exploring. Though inter-concept correlations have been utilized to address this issue recently, the very small number of instances in the minority classes often lead to the detection of imprecise correlations and unsatisfactory classification results. Meanwhile, correlation discovery is a computationally intensive task in the sense that it requires a deep analysis of very large and growing repositories. This dissertation further proposes a novel concept correlation analysis strategy framework that utilizes the correlations between the retrieval scores and labels. By integrating the correlation information, the proposed framework can help imbalanced data classification and enhance rare class (event or concept) mining even with trivial scores from the minority classes. Not only deep learning but also numerous other classification algorithms have been developed for a variety of data types. However, it is nearly impossible for one classifier to perform the best in all kinds of datasets all the time. Therefore, ensemble learning models which aim to take advantages of different classifiers have received a lot of attentions recently. In this dissertation, a scalable classifier ensemble framework assisted by a set of “judgers” is also proposed to integrate the outputs from multiple classifiers for multimedia big data classification. Specifically, based on the confusion matrices of different classifiers, a set of judgers are organized into a hierarchically structured decision model. A testing instance is first input to different classifiers, and then the classification results are passed to the proposed hierarchical structured decision model to derive the final result. The ensemble system can be run on Spark, which is designed for big data processing. All the proposed components are evaluated on multimedia datasets containing different kinds of data. The experimental results show the effectiveness of our framework in classifying severely imbalanced data with promising performance, and demonstrate that the proposed classifier ensemble framework outperforms several state-of-the-art model fusion approaches. Furthermore, the proposed framework is applied to two realworld applications, i.e., deep learning based text data analysis on an Amazon review dataset and efficient large-scale stance Analysis in Twitter, and achieves promising results in both. In additional, we also design a web-based information retrieval system and identify several future directions that could be explored to further improve the current work. To my dear parents and grandparents

We propose a multimedia analytics solution for getting insight into image collections by extending the powerful analytic capabilities of pivot tables, found in the ubiquitous spreadsheets, to multimedia. We formalize the concept of multimedia pivot tables and give design rules and methods for the multimodal summarization, structuring, and browsing of the collection based on these tables, all optimized to support an analyst in getting structural and conclusive insights. Our proposed solution provides truly interactive analytics on the visual content of image collections through concept detection results, as well as tags, geolocation, time, and other metadata. We have performed user experiments with novice users on a dataset from Flickr to improve the initial design and with expert users in marketing and multimedia analysis on two domain-specific datasets collected from Instagram. The results show that analysts are indeed capable of deriving structural and conclusive insights using the proposed multimedia analytics solution. On our website, videos of the system in action are available.

We propose a multimedia analytics solution for getting insight in image collections by extending the powerful method of pivot tables, found in the ubiquitous spreadsheets, to multimedia. Our proposed solution is designed by considering the characteristics of multimedia data as well as insight and provides integral access to visual content through concept detection results, tags, geolocation, and other metadata. We present a set of scenarios of using the pivot tables for a collection of images, tags, and metadata from Flickr. User experiments have been instrumental in realizing the final design presented in this paper. The accompanying video shows the solution in action.

We propose a few-shot adaptation framework, which bridges zero-shot learning and supervised many-shot learning, for semantic indexing of image and video data. Few-shot adaptation provides robust parameter estimation with few training examples, by optimizing the parameters of zero-shot learning and supervised many-shot learning simultaneously. In this method, first we build a zero-shot detector, and then update it by using the few examples. Our experiments show the effectiveness of the proposed framework on three datasets: TRECVID Semantic Indexing 2010, 2014, and ImageNET. On the ImageNET dataset, we show that our method outperforms recent few-shot learning methods. On the TRECVID 2014 dataset, we achieve 15.19~% and 35.98~% in Mean Average Precision under the zero-shot condition and the supervised condition, respectively. To the best of our knowledge, these are the best results on this dataset.

Video semantic indexing, which aims to detect objects, actions and scenes from video data, is one of important research topics in multimedia information processing. In the Text Retrieval Conference Video Retrieval Evaluation (TRECVID) workshop, many fundamental techniques for video processing have been developed and have been shown to be effective for real data such as Internet videos. They include extensions of deep learning techniques and image recognition techniques such as bag of visual words to video data. This paper reviews TRECVID activities with these techniques for semantic indexing. We also show the TokyoTech system using Gaussian-mixture-model (GMM) supervectors and deep convolutional neural networks (CNNs) with its experimental evaluation at TRECVID 2014.

Today's ubiquity of visual content as driven by the availability of broadband Internet, low-priced storage, and the omnipresence of camera equipped mobile devices conveys much of our thinking and feeling as individuals and as a society. As a result the growth of video repositories is increasing at enourmous rates with content now being embedded and shared through social media. To make use of this new form of social multimedia, concept detection, the automatic mapping of semantic concepts and video content has to be extended such that concept vocabularies are synchronized with current real-world events, systems can perform scalable concept learning with thousands of concepts, and high-level information such as sentiment can be extracted from visual content. To catch up with these demands the following three contributions are made in this thesis: (i) concept detection is linked to trending topics, (ii) visual learning from web videos is presented including the proper treatment of tags as concept labels, and (iii) the extension of concept detection with adjective noun pairs for sentiment analysis is proposed. In order for concept detection to satisfy users' current information needs, the notion of fixed concept vocabularies has to be reconsidered. This thesis presents a novel concept learning approach built upon dynamic vocabularies, which are automatically augmented with trending topics mined from social media. Once discovered, trending topics are evaluated by forecasting their future progression to predict high impact topics, which are then either mapped to an available static concept vocabulary or trained as individual concept detectors on demand. It is demonstrated in experiments on YouTube video clips that by a visual learning of trending topics, improvements of over 100% in concept detection accuracy can be achieved over static vocabularies (n=78,000). To remove manual efforts related to training data retrieval from YouTube and noise caused by tags being coarse, subjective and context-depedent, this thesis suggests an automatic concept-to-query mapping for the retrieval of relevant training video material, and active relevance filtering to generate reliable annotations from web video tags. Here, the relevance of web tags is modeled as a latent variable, which is combined with an active learning label refinement. In experiments on YouTube, active relevance filtering is found to outperform both automatic filtering and active learning approaches, leading to a reduction of required label inspections by 75% as compared to an expert annotated training dataset (n=100,000). Finally, it is demonstrated, that concept detection can serve as a key component to infer the sentiment reflected in visual content. To extend concept detection for sentiment analysis, adjective noun pairs (ANP) as novel entities for concept learning are proposed in this thesis. First a large-scale visual sentiment ontology consisting of 3,000 ANPs is automatically constructed by mining the web. From this ontology a mid-level representation of visual content – SentiBank – is trained to encode the visual presence of 1,200 ANPs. This novel approach of visual learning is validated in three independent experiments on sentiment prediction (n=2,000), emotion detection (n=807) and pornographic filtering (n=40,000). SentiBank is shown to outperform known low-level feature representations (sentiment prediction, pornography detection) or perform comparable to state-of-the art methods (emotion detection). Altogether, these contributions extend state-of-the-art concept detection approaches such that concept learning can be done autonomously from web videos on a large-scale, and can cope with novel semantic structures such as trending topics or adjective noun pairs, adding a new dimension to the understanding of video content.

This paper proposes an elastic spatial verification method for Instance Search, particularly for dealing with non-planar and non-rigid queries exhibiting complex spatial transformations. Different from existing models that map keypoints between images based on a linear transformation (e.g., affine, homography), our model exploits the topological arrangement of keypoints to address the non-linear spatial transformations that are extremely common in real life situations. In particular, we propose a novel technique to elastically verify the topological spatial consistency with the triangulated graph through a “sketch-and-match” scheme. The spatial topology configuration, emphasizing relative positioning rather than absolute coordinates, is first sketched by a triangulated graph, whose edges essentially capture the topological layout of the corresponding keypoints. Next, the spatial consistency is efficiently estimated as the number of common edges between the triangulated graphs. Compared to the existing methods, our technique is much more effective in modeling the complex spatial transformations of non-planar and non-rigid instances, while being compatible to instances with simple linear transformations. Moreover, our method is by nature more robust in spatial verification by considering the locations, rather than the local geometry of keypoints, which are sensitive to motions and viewpoint changes. We evaluate our method extensively on three years of TRECVID datasets, as well as our own dataset MQA, showing large improvement over other methods for the task of Instance Search.

This paper addresses the fundamental question -- How do humans recognize complex events in videos? Normally, humans view videos in a sequential manner. We hypothesize that humans can make high-level inference such as an event is present or not in a video, by looking at a very small number of frames not necessarily in a linear order. We attempt to verify this cognitive capability of humans and to discover the Minimally Needed Evidence (MNE) for each event. To this end, we introduce an online game based event quiz facilitating selection of minimal evidence required by humans to judge the presence or absence of a complex event in an open source video. Each video is divided into a set of temporally coherent microshots (1.5 secs in length) which are revealed only on player request. The player's task is to identify the positive and negative occurrences of the given target event with minimal number of requests to reveal evidence. Incentives are given to players for correct identification with the minimal number of requests. Our extensive human study using the game quiz validates our hypothesis - 55% of videos need only one microshot for correct human judgment and events of varying complexity require different amounts of evidence for human judgment. In addition, the proposed notion of MNE enables us to select discriminative features, drastically improving speed and accuracy of a video retrieval system.

The subject of this thesis is about image and video representations for visual recognition. This thesis first focuses on image search, both for image and textual queries, and then considers the classification and the localization of actions in videos. In image retrieval, images similar to the query image are retrieved from a large dataset. On this front, we propose an asymmetric version of the Hamming Embedding method, where the comparison of query and database descriptors relies on a vector-to-binary code comparison. For image classification, where the task is to identify if an image contains any instance of the queried category, we propose a novel approach based on a match kernel between images, more specifically based on Hamming Embedding similarity. We also present an effective variant of the SIFT descriptor, which leads to a better classification accuracy. Action classification is improved by several methods to better employ the motion inherent to videos. This is done by dominant motion compensation, and by introducing a novel descriptor based on kinematic features of the visual flow. The last contribution is devoted to action localization, whose objective is to determine where and when the action of interest appears in the video. A selective sampling strategy produces 2D+t sequences of bounding boxes, which drastically reduces the candidate locations. The method advantageously exploits a criterion that takes in account how motion related to actions deviates from the background motion. We thoroughly evaluated all the proposed methods on real world images and videos from challenging benchmarks. Our methods outperform the previously published related state of the art and remains competitive with the subsequently proposed methods.

The rise of the social media and video streaming industry provided us a plethora of videos and their corresponding descriptive information in the form of concepts (words) and textual video captions. Due to the mass amount of available videos and the textual data, today is the best time ever to study the Computer Vision and Machine Learning problems related to videos and text. In this dissertation, we tackle multiple problems associated with the joint understanding of videos and text. We first address the task of multi-concept video retrieval, where the input is a set of words as concepts, and the output is a ranked list of full-length videos. This approach deals with multiconcept input and prolonged length of videos by incorporating multi-latent variables to tie the information within each shot (short clip of a full-video) and across shots. Secondly, we address the problem of video question answering, in which, the task is to answer a question, in the form of Fill-In-the-Blank (FIB), given a video. Answering a question is a task of retrieving a word from a dictionary (all possible words suitable for an answer) based on the input question and video. Following the FIB problem, we introduce a new problem, called Visual Text Correction (VTC), i.e., detecting and replacing an inaccurate word in the textual description of a video. We propose a deep network that can simultaneously detect an inaccuracy in a sentence while benefiting 1D-CNNs/LSTMs to encode short/long term dependencies, and fix it by replacing the inaccurate word(s). Finally, as the last part of the dissertation, we propose to tackle the problem of video generation using user input natural language sentences. Our proposed video generation method constructs two distributions out of the input text, corresponding to the first and last frames latent representations. We generate high-fidelity videos by interpolating latent representations and a sequence of CNN based up-pooling blocks.

The goal of this paper is to recognize actions in video without the need for examples. Different from traditional zero-shot approaches we do not demand the design and specification of attribute classifiers and class-to-attribute mappings to allow for transfer from seen classes to unseen classes. Our key contribution is objects2action, a semantic word embedding that is spanned by a skip-gram model of thousands of object categories. Action labels are assigned to an object encoding of unseen video based on a convex combination of action and object affinities. Our semantic embedding has three main characteristics to accommodate for the specifics of actions. First, we propose a mechanism to exploit multiple-word descriptions of actions and objects. Second, we incorporate the automated selection of the most responsive objects per action. And finally, we demonstrate how to extend our zero-shot approach to the spatio-temporal localization of actions in video. Experiments on four action datasets demonstrate the potential of our approach.

The goal of this paper is to recognize actions in video without the need for examples. Different from traditional zero-shot approaches we do not demand the design and specification of attribute classifiers and class-to-attribute mappings to allow for transfer from seen classes to unseen classes. Our key contribution is objects2action, a semantic word embedding that is spanned by a skip-gram model of thousands of object categories. Action labels are assigned to an object encoding of unseen video based on a convex combination of action and object affinities. Our semantic embedding has three main characteristics to accommodate for the specifics of actions. First, we propose a mechanism to exploit multiple-word descriptions of actions and objects. Second, we incorporate the automated selection of the most responsive objects per action. And finally, we demonstrate how to extend our zero-shot approach to the spatio-temporal localization of actions in video. Experiments on four action datasets demonstrate the potential of our approach.

The classification of imbalanced datasets has recently attracted significant attention due to its implications in several real-world use cases. The classifiers developed on datasets with skewed distributions tend to favor the majority classes and are biased against the minority class. Despite extensive research interests, imbalanced data classification remains a challenge in data mining research, especially for multimedia data. Our attempt to overcome this hurdle is to develop a convolutional neural network CNN based deep learning solution integrated with a bootstrapping technique. Considering that convolutional neural networks are very computationally expensive coupled with big training datasets, we propose to extract features from pre-trained convolutional neural network models and feed those features to another full connected neutral network. Spark implementation shows promising performance of our model in handling big datasets with respect to feasibility and scalability.

The Internet has been witnessing an explosion of video content. According to a Cisco study, video content is estimated to account for 80% of all the entire world's internet traffic by 2019. Video data are becoming one of the most valuable sources to assess information and knowledge. However, existing video search solutions are still based on text matching (text-to-text search), and could fail for the huge volumes of videos that have little relevant metadata or no metadata at all. The need for large-scale and intelligent video search, which bridges the gap between the user's information need and the video content, seems to be urgent. In this thesis, we propose an accurate, efficient and scalable search method for video content. As opposed to text matching, the proposed method relies on automatic video content understanding, and allows for intelligent and flexible search paradigms over the video content, including text-to-video and text&video-to-video search. Suppose our goal is to search the videos about birthday party. In traditional text-to-text queries, we have to search the keywords in the user-generated metadata (titles or descriptions). In a text-to-video query, however, we might look for visual clues in the video content such as "cake", "gift" and "kids", audio clues like "birthday song" and "cheering sound", or visible text like "happy birthday". Text-to-video queries are flexible and can be further refined by Boolean and temporal operators. After watching the retrieved videos, the user may select a few interesting videos to find more relevant videos like these. This can be achieved by issuing a text&video-to-video query which adds the selected video examples to the query. The proposed method provides a new dimension of looking at content-based video search, from finding a simple concept like "puppy" to searching a complex incident like "a scene in urban area where people running away after an explosion". To achieve this ambitious goal, we propose several novel methods focusing on accuracy, efficiency and scalability in the novel search paradigm. First, we introduce a novel self-paced curriculum learning theory that allows for training more accurate semantic concepts. Second, we propose a novel and scalable approach to index semantic concepts that can significantly improve the search efficiency with minimum accuracy loss. Third, we design a novel video reranking algorithm that can boost accuracy for video retrieval. The extensive experiments demonstrate that the proposed methods are able to surpass state-of-the-art accuracy on multiple datasets. In addition, our method can efficiently scale up the search to hundreds of millions videos, and only takes about 0.2 second to search a semantic query on a collection of 100 million videos, 1 second to process a hybrid query over 1 million videos. Based on the proposed methods, we implement E-Lamp Lite, the first of its kind large-scale semantic search engine for Internet videos. According to National Institute of Standards and Technology (NIST), it achieved the best accuracy in the TRECVID Multimedia Event Detection (MED) 2013, 2014 and 2015, the most representative task for content-based video search. To the best of our knowledge, E-Lamp Lite is the first content-based semantic search engine that is capable of indexing and searching a collection of 100 million videos.

Techniques for recognizing high-level events in consumer videos on the Internet have many applications. Systems that produced state-of-the-art recognition performance usually contain modules requiring extensive computation, such as the extraction of the temporal motion trajectories, which cannot be deployed on large-scale datasets. In this paper, we provide a comprehensive study on efficient methods in this area and identify technical options for super fast event recognition in Internet videos. We start from analyzing a multimodal baseline that has produced good performance on popular benchmarks, by systematically evaluating each component in terms of both computational cost and contribution to recognition accuracy. After that, we identify alternative features, classifiers, and fusion strategies that can all be efficiently computed. In addition, we also provide a study on the following interesting question: for event recognition in Internet videos, what is the minimum number of visual and audio frames needed to obtain a comparable accuracy to that of using all the frames? Results on two rigorously designed datasets indicate that similar results can be maintained by using only a small portion of the visual frames. We also find that, different from the visual frames, the soundtracks contain little redundant information and thus sampling is always harmful. Integrating all the findings, our suggested recognition system is 2,350-fold faster than a baseline approach with even higher recognition accuracies. It recognizes 20 classes on a 120-second video sequence in just 1.78 seconds, using a regular desktop computer.

Supervised learning using deep convolutional neural network has shown its promise in large-scale image classification task. As a building block, it is now well positioned to be part of a larger system that tackles real-life multimedia tasks. An unresolved issue is that such model is trained on a static snapshot of data. Instead, this paper positions the training as a continuous learning process as new classes of data arrive. A system with such capability is useful in practical scenarios, as it gradually expands its capacity to predict increasing number of new classes. It is also our attempt to address the more fundamental issue: a good learning system must deal with new knowledge that it is exposed to, much as how human do. We developed a training algorithm that grows a network not only incrementally but also hierarchically. Classes are grouped according to similarities, and self-organized into levels. The newly added capacities are divided into component models that predict coarse-grained superclasses and those return final prediction within a superclass. Importantly, all models are cloned from existing ones and can be trained in parallel. These models inherit features from existing ones and thus further speed up the learning. Our experiment points out advantages of this approach, and also yields a few important open questions.

Searching for matches to high-dimensional vectors using hard/soft vector quantization is the most computationally expensive part of various computer vision algorithms including the bag of visual word (BoW). This paper proposes a fast computation method, Neighbor-to-Neighbor (NTN) search [1], which skips some calculations based on the similarity of input vectors. For example, in image classification using dense SIFT descriptors, the NTN search seeks similar descriptors from a point on a grid to an adjacent point. Applications of the NTN search to vector quantization, a Gaussian mixture model, sparse coding, and a kernel codebook for extracting image or video representation are presented in this paper. We evaluated the proposed method on image and video benchmarks: the PASCAL VOC 2007 Classification Challenge and the TRECVID 2010 Semantic Indexing Task. NTN-VQ reduced the coding cost by 77.4 percent, and NTN-GMM reduced it by 89.3 percent, without any significant degradation in classification performance.

Online discussion forums provide open workspace allowing users to share information, exchange ideas, address problems, and form groups. These forums feature multimodal posts and analyzing them requires a framework that can integrate heterogeneous information extracted from the posts, i.e. text, visual content and the information about user interactions with the online platform and each other. In this paper, we develop a generic framework that can be trained to identify communication behavior and patterns in relation to an entity of interest, be it user, image or text in internet forums. As the case study we use the analysis of violent online political extremism content, which has been a major challenge for domain experts. We demonstrate the generalizability and flexibility of our framework in predicting relational information between multimodal entities by conducting extensive experimentation around four practical use cases.

OF THE DISSERTATION INTEGRATING DEEP LEARNING WITH CORRELATION-BASED MULTIMEDIA SEMANTIC CONCEPT DETECTION by Hsin-Yu Ha Florida International University, 2015 Miami, Florida Professor Shu-Ching Chen, Major Professor The rapid advances in technologies make the explosive growth of multimedia data possible and available to the public. Multimedia data can be defined as data collection, which is composed of various data types and different representations. Due to the fact that multimedia data carries knowledgeable information, it has been widely adopted to different genera, like surveillance event detection, medical abnormality detection, and many others. To fulfill various requirements for different applications, it is important to effectively classify multimedia data into semantic concepts across multiple domains. In this dissertation, a correlation-based multimedia semantic concept detection framework is seamlessly integrated with the deep learning technique. The framework aims to explore implicit and explicit correlations among features and concepts while adopting different Convolutional Neural Network (CNN) architectures accordingly. First, the Feature Correlation Maximum Spanning Tree (FC-MST) is proposed to remove the redundant and irrelevant features based on the correlations between the features and positive concepts. FC-MST identifies the effective features and decides the initial layer’s dimension in CNNs. Second, the Negative-based Sampling method is proposed to alleviate the data imbalance issue by keeping only the representative negative instances in the training process. To adjust different sizes of training data, the number of iterations for the CNN

Nowadays, concept detection from multimedia data is considered as an emerging topic due to its applicability to various applications in both academia and industry. However, there are some inevitable challenges including the high volume and variety of multimedia data as well as its skewed distribution. To cope with these challenges, in this paper, a novel framework is proposed to integrate two correlation-based methods, Feature-Correlation Maximum Spanning Tree (FC-MST) and Negative-based Sampling (NS), with a well-known deep learning algorithm called Convolutional Neural Network (CNN). First, FC-MST is introduced to select the most relevant low-level features, which are extracted from multiple modalities, and to decide the input layer dimension of the CNN. Second, NS is adopted to improve the batch sampling in the CNN. Using NUS-WIDE image data set as a web-based application, the experimental results demonstrate the effectiveness of the proposed framework for semantic concept detection, comparing to other well-known classifiers.