Datasets, features, learning, and models in visual recognition

Visual recognition is a fundamental research topic in computer vision. This dissertation explores datasets, features, learning, and models used for visual recognition. In order to train visual models and evaluate different recognition algorithms, this dissertation develops an approach to collect object image datasets on web pages using an analysis of text around the image and of image appearance. This method exploits established online knowledge resources (Wikipedia pages for text; Flickr and Caltech data sets for images). The resources provide rich text and object appearance information. This dissertation describes results on two datasets. The first is Berg’s collection of 10 animal categories; on this dataset, we significantly outperform previous approaches. On an additional set of 5 categories, experimental results show the effectiveness of the method. Images are represented as features for visual recognition. This dissertation introduces a text-based image feature and demonstrates that it consistently improves performance on hard object classification problems. The feature is built using an auxiliary dataset of images annotated with tags, downloaded from the Internet. Image tags are noisy. The method obtains the text features of an unannotated image from the tags of its k-nearest neighbors in this auxiliary collection. A visual classifier presented with an object viewed under novel circumstances (say, a new viewing direction) must rely on its visual examples. This text feature may not change, because the auxiliary dataset likely contains a similar picture. While the tags associated with images are noisy, they are more stable when appearance changes. The performance of this feature is tested using PASCAL VOC 2006 and 2007 datasets. This feature performs well; it consistently improves the performance of visual object classifiers, and is particularly effective when the training dataset is small. With more and more collected training data, computational cost becomes a bottleneck, especially when training sophisticated classifiers such as kernelized SVM. This dissertation proposes a fast training algorithm called Stochastic Intersection Kernel Machine (SIKMA). This proposed training method will be useful for many vision problems, as it can produce a kernel classifier that is more accurate than a linear classifier, and can be trained on tens of thousands of examples in two minutes. It processes training examples one by one in a sequence, so memory cost is no longer the bottleneck to process large scale datasets. This dissertation applies this approach to train classifiers of Flickr groups with many group training examples. The resulting Flickr group prediction scores can be used to measure image similarity between two images. Experimental results on the Corel dataset and a PASCAL VOC dataset show the learned Flickr features perform better on image matching, retrieval, and classification than conventional visual features. Visual models are usually trained to best separate positive and negative training examples. However, when recognizing a large number of object categories, there may not be enough training examples for most objects, due to the intrinsic long-tailed distribution of objects in the real world. This dissertation proposes an approach to use comparative object similarity. The key insight is that, given a set of object categories which are similar and a set of categories which are dissimilar, a good object model should respond more strongly to examples from similar categories than to examples from dissimilar categories. This dissertation develops a regularized kernel machine algorithm to use this category dependent similarity regularization. Experiments on hundreds of categories show that our method can make significant improvement for categories with few or even no positive examples.

Context Models for Understanding Images and Videos

A computer vision system that has to interact in natural language needs to understand the visual appearance of interactions between objects along with the appearance of objects themselves. Relationships between objects are frequently mentioned in queries of tasks like semantic image retrieval, image captioning, visual question answering and natural language object detection. Hence, it is essential to model context between objects for solving these tasks. In the first part of this thesis, we present a technique for detecting an object mentioned in a natural language query. Specifically, we work with referring expressions which are sentences that identify a particular object instance in an image. In many referring expressions, an object is described in relation to another object using prepositions, comparative adjectives, action verbs etc. Our proposed technique can identify both the referred object and the context object mentioned in such expressions. Context is also useful for incrementally understanding scenes and videos. In the second part of this thesis, we propose techniques for searching for objects in an image and events in a video. Our proposed incremental algorithms use the context from previously explored regions to prioritize the regions to explore next. The advantage of incremental understanding is restricting the amount of computation time and/or resources spent for various detection tasks. Our first proposed technique shows how to learn context in indoor scenes in an implicit manner and use it for searching for objects. The second technique shows how explicitly written context rules of one-on-one basketball can be used to sequentially detect events in a game.

Scene Mapping and Understanding by Robotic Vision

The first mechanical Automaton concept was found in a Chinese text written in the 3rd century BC, while Computer Vision was born in the late 1960s. Therefore, visual perception applied to machines (i.e. the Machine Vision) is a young and exciting alliance. When robots came in, the new field of Robotic Vision was born, and these terms began to be erroneously interchanged. In short, we can say that Machine Vision is an engineering domain, which concern the industrial use of Vision. The Robotic Vision, instead, is a research field that tries to incorporate robotics aspects in computer vision algorithms. Visual Servoing, for example, is one of the problems that cannot be solved by computer vision only. Accordingly, a large part of this work deals with boosting popular Computer Vision techniques by exploiting robotics: e.g. the use of kinematics to localize a vision sensor, mounted as the robot end-effector. The remainder of this work is dedicated to the counterparty, i.e. the use of computer vision to solve real robotic problems like grasping objects or navigate avoiding obstacles. Will be presented a brief survey about mapping data structures most widely used in robotics along with SkiMap, a novel sparse data structure created both for robotic mapping and as a general purpose 3D spatial index. Thus, several approaches to implement Object Detection and Manipulation, by exploiting the aforementioned mapping strategies, will be proposed, along with a completely new Machine Teaching facility in order to simply the training procedure of modern Deep Learning networks.

Robotic Sensing and Manipulation of Deformable Linear Objects with Learning-based methods

Nowadays robotic applications are widespread and most of the manipulation tasks are efficiently solved. However, Deformable-Objects (DOs) still represent a huge limitation for robots. The main difficulty in DOs manipulation is dealing with the shape and dynamics uncertainties, which prevents the use of model-based approaches (since they are excessively computationally complex) and makes sensory data difficult to interpret. This thesis reports the research activities aimed to address some applications in robotic manipulation and sensing of Deformable-Linear-Objects (DLOs), with particular focus to electric wires. In all the works, a significant effort was made in the study of an effective strategy for analyzing sensory signals with various machine learning algorithms. In the former part of the document, the main focus concerns the wire terminals, i.e. detection, grasping, and insertion. First, a pipeline that integrates vision and tactile sensing is developed, then further improvements are proposed for each module. A novel procedure is proposed to gather and label massive amounts of training images for object detection with minimal human intervention. Together with this strategy, we extend a generic object detector based on Convolutional-Neural-Networks for orientation prediction. The insertion task is also extended by developing a closed-loop control capable to guide the insertion of a longer and curved segment of wire through a hole, where the contact forces are estimated by means of a Recurrent-Neural-Network. In the latter part of the thesis, the interest shifts to the DLO shape. Robotic reshaping of a DLO is addressed by means of a sequence of pick-and-place primitives, while a decision making process driven by visual data learns the optimal grasping locations exploiting Deep Q-learning and finds the best releasing point. The success of the solution leverages on a reliable interpretation of the DLO shape. For this reason, further developments are made on the visual segmentation.

Production Lot Detection through Synthetic Generated Data

Correctness of information gathered in production environments is an essential part of quality assurance processes in many industries, this task is often performed by human resources who visually take annotations in various steps of the production flow. Depending on the performed task the correlation between where exactly the information is gathered and what it represents is more than often lost in the process. The lack of labeled data places a great boundary on the application of deep neural networks aimed at object detection tasks, moreover supervised training of deep models requires a great amount of data to be available. Reaching an adequate large collection of labeled images through classic techniques of data annotations is an exhausting and costly task to perform, not always suitable for every scenario. A possible solution is to generate synthetic data that replicates the real one and use it to fine-tune a deep neural network trained on one or more source domains to a different target domain. The purpose of this thesis is to show a real case scenario where the provided data were both in great scarcity and missing the required annotations. Sequentially a possible approach is presented where synthetic data has been generated to address those issues while standing as a training base of deep neural networks for object detection, capable of working on images taken in production-like environments. Lastly, it compares performance on different types of synthetic data and convolutional neural networks used as backbones for the model.

Detecção e seguimento de objectos em imagens termográficas: análise experimental de modelos de descrição

A instalação de sistemas de videovigilância, no interior ou exterior, em locais como aeroportos, centros comerciais, escritórios, edifícios estatais, bases militares ou casas privadas tem o intuito de auxiliar na tarefa de monitorização do local contra eventuais intrusos. Com estes sistemas é possível realizar a detecção e o seguimento das pessoas que se encontram no ambiente local, tornando a monitorização mais eficiente. Neste contexto, as imagens típicas (imagem natural e imagem infravermelha) são utilizadas para extrair informação dos objectos detectados e que irão ser seguidos. Contudo, as imagens convencionais são afectadas por condições ambientais adversas como o nível de luminosidade existente no local (luzes muito fortes ou escuridão total), a presença de chuva, de nevoeiro ou de fumo que dificultam a tarefa de monitorização das pessoas. Deste modo, tornou‐se necessário realizar estudos e apresentar soluções que aumentem a eficácia dos sistemas de videovigilância quando sujeitos a condições ambientais adversas, ou seja, em ambientes não controlados, sendo uma das soluções a utilização de imagens termográficas nos sistemas de videovigilância. Neste documento são apresentadas algumas das características das câmaras e imagens termográficas, assim como uma caracterização de cenários de vigilância. Em seguida, são apresentados resultados provenientes de um algoritmo que permite realizar a segmentação de pessoas utilizando imagens termográficas. O maior foco desta dissertação foi na análise dos modelos de descrição (Histograma de Cor, HOG, SIFT, SURF) para determinar o desempenho dos modelos em três casos: distinguir entre uma pessoa e um carro; distinguir entre duas pessoas distintas e determinar que é a mesma pessoa ao longo de uma sequência. De uma forma sucinta pretendeu‐se, com este estudo, contribuir para uma melhoria dos algoritmos de detecção e seguimento de objectos em sequências de vídeo de imagens termográficas. No final, através de uma análise dos resultados provenientes dos modelos de descrição, serão retiradas conclusões que servirão de indicação sobre qual o modelo que melhor permite discriminar entre objectos nas imagens termográficas.

GPU implementation of the simplex identification via split augmented Lagrangian

Hyperspectral imaging can be used for object detection and for discriminating between different objects based on their spectral characteristics. One of the main problems of hyperspectral data analysis is the presence of mixed pixels, due to the low spatial resolution of such images. This means that several spectrally pure signatures (endmembers) are combined into the same mixed pixel. Linear spectral unmixing follows an unsupervised approach which aims at inferring pure spectral signatures and their material fractions at each pixel of the scene. The huge data volumes acquired by such sensors put stringent requirements on processing and unmixing methods. This paper proposes an efficient implementation of a unsupervised linear unmixing method on GPUs using CUDA. The method finds the smallest simplex by solving a sequence of nonsmooth convex subproblems using variable splitting to obtain a constraint formulation, and then applying an augmented Lagrangian technique. The parallel implementation of SISAL presented in this work exploits the GPU architecture at low level, using shared memory and coalesced accesses to memory. The results herein presented indicate that the GPU implementation can significantly accelerate the method's execution over big datasets while maintaining the methods accuracy.

Parallel GPU architecture for hyperspectral unmixing based on augmented Lagrangian method

Hyperspectral imaging has become one of the main topics in remote sensing applications, which comprise hundreds of spectral bands at different (almost contiguous) wavelength channels over the same area generating large data volumes comprising several GBs per flight. This high spectral resolution can be used for object detection and for discriminate between different objects based on their spectral characteristics. One of the main problems involved in hyperspectral analysis is the presence of mixed pixels, which arise when the spacial resolution of the sensor is not able to separate spectrally distinct materials. Spectral unmixing is one of the most important task for hyperspectral data exploitation. However, the unmixing algorithms can be computationally very expensive, and even high power consuming, which compromises the use in applications under on-board constraints. In recent years, graphics processing units (GPUs) have evolved into highly parallel and programmable systems. Specifically, several hyperspectral imaging algorithms have shown to be able to benefit from this hardware taking advantage of the extremely high floating-point processing performance, compact size, huge memory bandwidth, and relatively low cost of these units, which make them appealing for onboard data processing. In this paper, we propose a parallel implementation of an augmented Lagragian based method for unsupervised hyperspectral linear unmixing on GPUs using CUDA. The method called simplex identification via split augmented Lagrangian (SISAL) aims to identify the endmembers of a scene, i.e., is able to unmix hyperspectral data sets in which the pure pixel assumption is violated. The efficient implementation of SISAL method presented in this work exploits the GPU architecture at low level, using shared memory and coalesced accesses to memory.

Groundtruth system for underwater benchmarking

Oceans - San Diego, 2013

Deteção Automática de Conteúdos Explícitos em Sequências de Vídeo

Nos últimos anos, o fácil acesso em termos de custos, ferramentas de produção, edição e distribuição de conteúdos audiovisuais, contribuíram para o aumento exponencial da produção diária deste tipo de conteúdos. Neste paradigma de superabundância de conteúdos multimédia existe uma grande percentagem de sequências de vídeo que contém material explícito, sendo necessário existir um controlo mais rigoroso, de modo a não ser facilmente acessível a menores. O conceito de conteúdo explícito pode ser caraterizado de diferentes formas, tendo o trabalho descrito neste documento incidido sobre a deteção automática de nudez feminina presente em sequências de vídeo. Este processo de deteção e classificação automática de material para adultos pode constituir uma ferramenta importante na gestão de um canal de televisão. Diariamente podem ser recebidas centenas de horas de material sendo impraticável a implementação de um processo manual de controlo de qualidade. A solução criada no contexto desta dissertação foi estudada e desenvolvida em torno de um produto especifico ligado à área do broadcasting. Este produto é o mxfSPEEDRAIL F1000, sendo este uma solução da empresa MOG Technologies. O objetivo principal do projeto é o desenvolvimento de uma biblioteca em C++, acessível durante o processo de ingest, que permita, através de uma análise baseada em funcionalidades de visão computacional, detetar e sinalizar na metadata do sinal, quais as frames que potencialmente apresentam conteúdo explícito. A solução desenvolvida utiliza um conjunto de técnicas do estado da arte adaptadas ao problema a tratar. Nestas incluem-se algoritmos para realizar a segmentação de pele e deteção de objetos em imagens. Por fim é efetuada uma análise critica à solução desenvolvida no âmbito desta dissertação de modo a que em futuros desenvolvimentos esta seja melhorada a nível do consumo de recursos durante a análise e a nível da sua taxa de sucesso.

Digital beamforming and moving target indication for future spaceborne synthetic aperture radar

During the last decade the interest on space-borne Synthetic Aperture Radars (SAR) for remote sensing applications has grown as testified by the number of recent and forthcoming missions as TerraSAR-X, RADARSAT-2, COSMO-kyMed, TanDEM-X and the Spanish SEOSAR/PAZ. In this sense, this thesis proposes to study and analyze the performance of the state-of-the-Art space-borne SAR systems, with modes able to provide Moving Target Indication capabilities (MTI), i.e. moving object detection and estimation. The research will focus on the MTI processing techniques as well as the architecture and/ or configuration of the SAR instrument, setting the limitations of the current systems with MTI capabilities, and proposing efficient solutions for the future missions. Two European projects, to which the Universitat Politècnica de Catalunya provides support, are an excellent framework for the research activities suggested in this thesis. NEWA project proposes a potential European space-borne radar system with MTI capabilities in order to fulfill the upcoming European security policies. This thesis will critically review the state-of-the-Art MTI processing techniques as well as the readiness and maturity level of the developed capabilities. For each one of the techniques a performance analysis will be carried out based on the available technologies, deriving a roadmap and identifying the different technological gaps. In line with this study a simulator tool will be developed in order to validate and evaluate different MTI techniques in the basis of a flexible space-borne radar configuration. The calibration of a SAR system is mandatory for the accurate formation of the SAR images and turns to be critical in the advanced operation modes as MTI. In this sense, the SEOSAR/PAZ project proposes the study and estimation of the radiometric budget. This thesis will also focus on an exhaustive analysis of the radiometric budget considering the current calibration concepts and their possible limitations. In the framework of this project a key point will be the study of the Dual Receive Antenna (DRA) mode, which provides MTI capabilities to the mission. An additional aspect under study is the applicability of the Digital Beamforming on multichannel and/or multistatic radar platforms, which conform potential solutions for the NEWA project with the aim to fully exploit its capability jointly with MTI techniques.

Event-related potential analyses via single-subject topographic classification.

Introduction: Responses to external stimuli are typically investigated by averaging peri-stimulus electroencephalography (EEG) epochs in order to derive event-related potentials (ERPs) across the electrode montage, under the assumption that signals that are related to the external stimulus are fixed in time across trials. We demonstrate the applicability of a single-trial model based on patterns of scalp topographies (De Lucia et al, 2007) that can be used for ERP analysis at the single-subject level. The model is able to classify new trials (or groups of trials) with minimal a priori hypotheses, using information derived from a training dataset. The features used for the classification (the topography of responses and their latency) can be neurophysiologically interpreted, because a difference in scalp topography indicates a different configuration of brain generators. An above chance classification accuracy on test datasets implicitly demonstrates the suitability of this model for EEG data. Methods: The data analyzed in this study were acquired from two separate visual evoked potential (VEP) experiments. The first entailed passive presentation of checkerboard stimuli to each of the four visual quadrants (hereafter, "Checkerboard Experiment") (Plomp et al, submitted). The second entailed active discrimination of novel versus repeated line drawings of common objects (hereafter, "Priming Experiment") (Murray et al, 2004). Four subjects per experiment were analyzed, using approx. 200 trials per experimental condition. These trials were randomly separated in training (90%) and testing (10%) datasets in 10 independent shuffles. In order to perform the ERP analysis we estimated the statistical distribution of voltage topographies by a Mixture of Gaussians (MofGs), which reduces our original dataset to a small number of representative voltage topographies. We then evaluated statistically the degree of presence of these template maps across trials and whether and when this was different across experimental conditions. Based on these differences, single-trials or sets of a few single-trials were classified as belonging to one or the other experimental condition. Classification performance was assessed using the Receiver Operating Characteristic (ROC) curve. Results: For the Checkerboard Experiment contrasts entailed left vs. right visual field presentations for upper and lower quadrants, separately. The average posterior probabilities, indicating the presence of the computed template maps in time and across trials revealed significant differences starting at ~60-70 ms post-stimulus. The average ROC curve area across all four subjects was 0.80 and 0.85 for upper and lower quadrants, respectively and was in all cases significantly higher than chance (unpaired t-test, p<0.0001). In the Priming Experiment, we contrasted initial versus repeated presentations of visual object stimuli. Their posterior probabilities revealed significant differences, which started at 250ms post-stimulus onset. The classification accuracy rates with single-trial test data were at chance level. We therefore considered sub-averages based on five single trials. We found that for three out of four subjects' classification rates were significantly above chance level (unpaired t-test, p<0.0001). Conclusions: The main advantage of the present approach is that it is based on topographic features that are readily interpretable along neurophysiologic lines. As these maps were previously normalized by the overall strength of the field potential on the scalp, a change in their presence across trials and between conditions forcibly reflects a change in the underlying generator configurations. The temporal periods of statistical difference between conditions were estimated for each training dataset for ten shuffles of the data. Across the ten shuffles and in both experiments, we observed a high level of consistency in the temporal periods over which the two conditions differed. With this method we are able to analyze ERPs at the single-subject level providing a novel tool to compare normal electrophysiological responses versus single cases that cannot be considered part of any cohort of subjects. This aspect promises to have a strong impact on both basic and clinical research.

Electrical neuroimaging of memory discrimination based on single-trial multisensory learning.

Multisensory experiences influence subsequent memory performance and brain responses. Studies have thus far concentrated on semantically congruent pairings, leaving unresolved the influence of stimulus pairing and memory sub-types. Here, we paired images with unique, meaningless sounds during a continuous recognition task to determine if purely episodic, single-trial multisensory experiences can incidentally impact subsequent visual object discrimination. Psychophysics and electrical neuroimaging analyses of visual evoked potentials (VEPs) compared responses to repeated images either paired or not with a meaningless sound during initial encounters. Recognition accuracy was significantly impaired for images initially presented as multisensory pairs and could not be explained in terms of differential attention or transfer of effects from encoding to retrieval. VEP modulations occurred at 100-130ms and 270-310ms and stemmed from topographic differences indicative of network configuration changes within the brain. Distributed source estimations localized the earlier effect to regions of the right posterior temporal gyrus (STG) and the later effect to regions of the middle temporal gyrus (MTG). Responses in these regions were stronger for images previously encountered as multisensory pairs. Only the later effect correlated with performance such that greater MTG activity in response to repeated visual stimuli was linked with greater performance decrements. The present findings suggest that brain networks involved in this discrimination may critically depend on whether multisensory events facilitate or impair later visual memory performance. More generally, the data support models whereby effects of multisensory interactions persist to incidentally affect subsequent behavior as well as visual processing during its initial stages.

The brain uses single-trial multisensory memories to discriminate without awareness.

Multisensory experiences enhance perceptions and facilitate memory retrieval processes, even when only unisensory information is available for accessing such memories. Using fMRI, we identified human brain regions involved in discriminating visual stimuli according to past multisensory vs. unisensory experiences. Subjects performed a completely orthogonal task, discriminating repeated from initial image presentations intermixed within a continuous recognition task. Half of initial presentations were multisensory, and all repetitions were exclusively visual. Despite only single-trial exposures to initial image presentations, accuracy in indicating image repetitions was significantly improved by past auditory-visual multisensory experiences over images only encountered visually. Similarly, regions within the lateral-occipital complex-areas typically associated with visual object recognition processes-were more active to visual stimuli with multisensory than unisensory pasts. Additional differential responses were observed in the anterior cingulate and frontal cortices. Multisensory experiences are registered by the brain even when of no immediate behavioral relevance and can be used to categorize memories. These data reveal the functional efficacy of multisensory processing.