Modern computer vision has seen recently significant progress in learning visual concepts from examples. This progress has been fuelled by recent models of visual appearance as well as recently collected large-scale datasets of manually annotated still images. Video is a promising alternative, as it inherently contains much richer information compared to still images. For instance, in video we can observe an object move which allows us to differentiate it from its surroundings, or we can observe a smooth transition between different viewpoints of the same object instance. This richness in information allows us to effectively tackle tasks that would otherwise be very difficult if we only considered still images, or even adress tasks that are video-specific. Our first contribution is a computationally efficient technique for video object segmentation. Our method relies solely on motion in order to rapidly create a rough initial estimate of the foreground object. This rough initial estimate is then refined through an energy formulation to be spatio-temporally smooth. The method is able to handle rapidly moving backgrounds and objects, as well as non-rigid deformations and articulations without having prior knowledge about the objects appearance, size or location. In addition to this class-agnostic method, we present a class-specific method that incorporates additional class-specific appearance cues when the class of the foreground object is known in advance (e.g. a video of a car). For our second contribution, we propose a novel model for temporal video alignment with regard to the viewpoint of the foreground object (i.e., a pair of aligned frames shows the same object viewpoint) Our work relies on our video object segmentation technique to automatically localise the foreground objects and extract appearance measurements solely from them instead of the background. Our model is able to temporally align realistic videos, where events may occur in a different order, or occur only in one of the videos. This is in contrast to previous works that typically assume that the videos show a scripted sequence of events and can simply be aligned by stretching or compressing one of the videos. As a final contribution, we once again use our video object segmentation technique as a basis for automatic visual aspect discovery from videos of an object class. Compared to previous works, we use a broader definition of an aspect that considers four factors of variation: viewpoint, articulated pose, occlusions and cropping by the image border. We pose the aspect discovery task as a clustering problem and provide an extensive experimental exploration on the benefits of object segmentation for this task.

This paper confirms the existence of short-term and long-term variation of the required bandwidth for MPEG videostreams. We show how the use of a small amount of buffering and GOP grouping can significantly reduce the effect of the short-term variation. By introducing a number of bandwidth renegotiation techniques, which can be applied to MPEG video streams in general, we are able to reduce the effect of long-term variation. These techniques include those that need the a priori knowledge of frame sizes as well as one that can renegotiate dynamically. A costing algorithm has also been introduced in order to compare various proposals against each other.

Les codeurs vidéo état de l’art utilisent des transformées pour assurer une représentation compacte du signal. L’étape de transformation constitue le domaine dans lequel s’effectue la compression, pourtant peu de variabilité dans les types de transformations est constatée dans les systèmes de codage vidéo normalisés : souvent, une seule transformée est considérée, habituellement la transformée en cosinus discrète (DCT). Récemment, d’autres transformées ont commencé à être considérées en complément de la DCT. Par exemple, dans le dernier standard de compression vidéo, nommé HEVC (High Efficiency Video Coding), les blocs de taille 4x4 peuvent utiliser la transformée en sinus discrète (DST), de plus, il est également possible de ne pas les transformer. Ceci révèle un intérêt croissant pour considérer une pluralité de transformées afin d’augmenter les taux de compression. Cette thèse se concentre sur l’extension de HEVC au travers de l’utilisation de multiples transformées. Après une introduction générale au codage vidéo et au codage par transformée, une étude détaillée de deux méthodes de construction de transformations est menée : la transformée de Karhunen Loève (KLT) et une transformée optimisée en débit et distorsion sont considérées. Ces deux méthodes sont comparées entre-elles en substituant les transformées utilisées par HEVC. Une expérimentation valide la pertinence des approches. Un schéma de codage qui incorpore et augmente l’utilisation de multiples transformées est alors introduit : plusieurs transformées sont mises à disposition de l’encodeur, qui sélectionne celle qui apporte le meilleur compromis dans le plan débit distorsion. Pour ce faire, une méthode de construction qui permet de concevoir des systèmes comportant de multiples transformations est décrite. Avec ce schéma de codage, le débit est significativement réduit par rapport à HEVC, tout particulièrement lorsque les transformées sont nombreuses et complexes à mettre en oeuvre. Néanmoins, ces améliorations viennent au prix d’une complexité accrue en termes d’encodage, de décodage et de contrainte de stockage. En conséquence, des simplifications sont considérées dans la suite du document, qui ont vocation à limiter l’impact en réduction de débit. Une première approche est introduite dans laquelle des transformées incomplètes sont motivées. Les transformations de ce type utilisent un seul vecteur de base, et sont conçues pour travailler de concert avec les transformations de HEVC. Cette technique est évaluée et apporte une réduction de complexité significative par rapport au précédent système, bien que la réduction de débit soit modeste. Une méthode systématique, qui détermine les meilleurs compromis entre le nombre de transformées et l’économie de débit est alors définie. Cette méthode utilise deux types différents de transformée : basés sur des transformées orthogonales séparables et des transformées trigonométriques discrètes (DTT) en particulier. Plusieurs points d’opération sont présentés qui illustrent plusieurs compromis complexité / gain en débit. Ces systèmes révèlent l’intérêt de l’utilisation de transformations multiples pour le codage vidéo. / State of the art video codecs use transforms to ensure a compact signal representation. The transform stage is where compression takes place, however, little variety is observed in the type of transforms used for standardised video coding schemes: often, a single transform is considered, usually a Discrete Cosine Transform (DCT). Recently, other transforms have started being considered in addition to the DCT. For instance, in the latest video coding standard, High Efficiency Video Coding (HEVC), the 4x4 sized blocks can make use of the Discrete Sine Transform (DST) and, in addition, it also possible not to transform them. This fact reveals an increasing interest to consider a plurality of transforms to achieve higher compression rates. This thesis focuses on extending HEVC through the use of multiple transforms. After a general introduction to video compression and transform coding, two transform designs are studied in detail: the Karhunen Loève Transform (KLT) and a Rate-Distortion Optimised Transform are considered. These two methods are compared against each other by replacing the transforms in HEVC. This experiment validates the appropriateness of the design. A coding scheme that incorporates and boosts the use of multiple transforms is introduced: several transforms are made available to the encoder, which chooses the one that provides the best rate-distortion trade-off. Consequently, a design method for building systems using multiple transforms is also described. With this coding scheme, significant amounts of bit-rate savings are achieved over HEVC, especially when using many complex transforms. However, these improvements come at the expense of increased complexity in terms of coding, decoding and storage requirements. As a result, simplifications are considered while limiting the impact on bit-rate savings. A first approach is introduced, in which incomplete transforms are used. This kind of transforms use one single base vector and are conceived to work as companions of the HEVC transforms. This technique is evaluated and provides significant complexity reductions over the previous system, although the bit-rate savings are modest. A systematic method, which specifically determines the best trade-offs between the number of transforms and bit-rate savings, is designed. This method uses two different types of transforms based separable orthogonal transforms and Discrete Trigonometric Transforms (DTTs) in particular. Several designs are presented, allowing for different complexity and bitrate savings trade-offs. These systems reveal the interest of using multiple transforms for video coding.

There is a significant and previously unidentified ethics problem with many e-games; many of them are designed in such a way that they encourage and train game players to follow a narrow and dangerous model of decision making. It is argued that extending this model of decision making beyond an e-game's virtual reality has significant negative societal consequences. Unfortunately most e-game courses focus primarily on game engine design and other technical issues. E-game curricula and e-game designers need to follow standards which recognize this ethical concern and recognize that their work is not independent from the societal impacts of the technology they develop. Modifying design approaches can reduce the problems caused by the ethical decision making model.

Le domaine de la mesure des caractéristiques du trafic transitant sur un réseau a été largement traité par une vaste communauté de chercheurs, en premier lieu pour répondre aux attentes des opérateurs fournisseurs d'accès à Internet. En effet, leur première préoccupation est de savoir quel type de trafic ils doivent transporter. Une des principales caractéristiques de l'Internet est qu'il évolue très vite, de sorte que le besoin de mesures du trafic grand public ne se tarit jamais. Dans ce travail, nous abordons la question de la mesure du trafic Internet grand public par deux perspectives différentes: les mesures passives et les mesures actives. Dans la première partie de cette thèse, nous capturons et analysons passivement les statistiques des connections d'utilisateurs d'Internet durant plus d'une semaine. Nous utilisons ces données pour réviser et approfondir notre connaissance du trafic Internet résidentiel. Ensuite, nous utilisons des méthodes de regroupement pour créer des ensembles d'utilisateurs en fonctions des applications qu'ils utilisent. Nous apprenons donc qu'une vaste majorité des clients se connectent à Internet principalement pour surfer sur le Web et regarder des vidéos en streaming. Ces données nous servent aussi à évaluer de nouvelles possibilités de contrôler le trafic d'une plateforme ADSL. Comme la principale partie du trafic provient du vidéo streaming, nous prenons plusieurs instantanés de ce trafic avec des captures paquet durant une période de plusieurs années, ceci pour comprendre précisément l'évolution de ce trafic. De plus, nous analysons et relions la performance du vidéo streaming, définie par des indicateurs de qualité de service, au comportement des utilisateurs de ce service. Dans la deuxième partie de cette thèse, nous tirons parti de cette connaissance pour concevoir une sonde active capable de mesurer la qualité d'expérience des sites de vidéo streaming. Nous avons modélisé la lecture des vidéos \emph{streaming} pour pouvoir déterminer leur qualité telle qu'elle est perçue par les utilisateurs. Grâce à cet outil, nous pouvons comprendre l'impact de la sélection du serveur vidéo et du serveur DNS sur la perception de la qualité vidéo par l'utilisateur. De plus, la possibilité de réaliser des mesures depuis divers opérateurs, nous permet de détailler les politiques de distribution vidéo utilisées par les sites de streaming. / Network tomography is the study of a network's traffic characteristics using measures. This subject has already been addressed by a whole community of researchers, especially to answer the need for knowledge of residential Internet traffic that ISPs have to carry. One of the main aspects of the Internet is that it evolves very quickly, so that there is a never ending need for Internet measurements. In this work, we address the issue of residential Internet measure from two different perspectives: passive measurements and active measurements. In the first part of this thesis, we passively collect and analyse statistics of residential users' connections spanning over a whole week. We use this data to update and deepen our knowledge of Internet residential traffic. Then, we use clustering methods to form groups of users according to the application they use. This shows how the vast majority of customers are now using the Internet mainly for Web browsing and watching video Streaming. This data is also used to evaluate new opportunities for managing the traffic of a local ADSL platform. As the main part of the traffic is video streaming, we use multiple snapshots of packet captures of this traffic over a period of many years to accurately understand its evolution. Moreover we analyse and correlate its performance, defined out of quality of service indicators, to the behavior of the users of this service. In the second part of this thesis, we take advantage of this knowledge to design a new tool for actively probing the quality of experience of video streaming sites. We have modeled the playback of streaming videos so that we are able to figure out its quality as perceived by the users. With this tool, we can understand the impact of the video server selection and the DNS servers on the user's perception of the video quality. Moreover the ability to perform the experiments on different ISPs allows us to further dig into the delivery policies of video streaming sites.

This thesis focuses on the interaction between digital video quality and channel coding in a wireless communication system. Digital video is a high-bandwidth, computationally intensive application. The recent allocation of large tracks of spectrum by the FCC has made possible the design and implementation of personal wireless digital video devices for several applications, from personal communications to surveillance. Wireless video research has centered on contextual coding mechanisms; different sections of a video data stream are coded differently based on the perceived importance of the coded bits. Besides the complexity introduced into a system by this type of coding, no metric based solely on physical layer parameters can be used to predict video quality across different system architectures. A tool was built to explore the video/channel coding relationship. This tool simulates a packet-based digital wireless transmission in various noise and interference environments. The basic communications system models the DAVIC (Digital Audio-Visual Council) layout for the LMDS (Local Multipoint Distribution Service) system and includes several error control algorithms and a packetizing algorithm that is MPEG-compliant. This work introduces a statistical approach to monitoring the video quality. The error-event-mean-arrival-rate, lambda_distance, is a metric derived from the physical layer that can be used to predict the expected video quality across systems with different channel coding algorithms. This metric proved to be uniformly consistent in predicting video quality for sequences corrupted by Gaussian and non-Gaussian noise and protected by a variety of error correction techniques. / Master of Science

We show that information leakage occurs in video over IP traffic, including for encrypted payloads. It is possible to detect events occurring in the field of view of a camera streaming live video through analysis of network traffic metadata including arrival time between packets, packets sizes, and video stream bandwidth. Event detection through metadata analysis is possible even when common encryption techniques are applied to the video stream such as SSL or AES. We have observed information leakage across multiple codes and cameras. Through timestamps added to the x264 codec, we establish a basis for detectability of events via packet timing. Laboratory experiments confirm that this event detection is possible in practice and repeatable. By collecting network traffic captures from over 100 Skype video calls we are able to see the impact of this information leakage under a variety of conditions.

published_or_final_version / abstract / Electrical and Electronic Engineering / Master / Master of Philosophy

No description available.

No description available.