As a core problem in video analysis, action recognition is of great significance for many higher-level tasks, both in research and industrial applications. With more and more video data being produced and shared daily, effective automatic action recognition methods are needed. Although, many deep-learning methods have been proposed to solve the problem, recent research reveals that single-stream, RGB-based networks are always outperformed by two-stream networks using both RGB and optical flow as inputs. This dependence on optical flow, which indicates a deficiency in learning motion, is present not only in 2D networks but also in 3D networks. This is somewhat surprising since 3D networks are explicitly designed for spatio-temporal learning. In this thesis, we assume that this deficiency is caused by difficulties associated with learning from videos exhibiting strong temporal variations, such as sudden motion, occlusions, acceleration, or deceleration. Temporal variations occur commonly in real-world videos and force a neural network to account for them, but often are not useful for recognizing actions at coarse granularity. We propose a Dynamic Equilibrium Module (DEM) for spatio-temporal learning through adaptive Eulerian motion manipulation. The proposed module can be inserted into existing networks with separate spatial and temporal convolutions, like the R(2+1)D model, to effectively handle temporal video variations and learn more robust spatio-temporal features. We demonstrate performance gains due to the use of DEM in the R(2+1)D model on miniKinetics, UCF-101, and HMDB-51 datasets.

Can compression parameters used in video encoding be estimated, given only the visual information of the resulting compressed video? If so, these parameters could potentially improve existing parametric video quality estimation models. Today, parametric models use information like bitrate to estimate the quality of a given video. This method is inaccurate since it does not consider the coding complexity of a video. The constant rate factor (CRF) parameter for h.264 encoding aims to keep the quality constant while varying the bitrate, if the CRF for a video is known together with bitrate, a better quality estimate could potentially be achieved. In recent years, artificial neural networks and specifically convolutional neural networks have shown great promise in the field of image processing. In this thesis, convolutional neural networks are investigated as a way of estimating the constant rate factor parameter for a degraded video by identifying the compression artifacts and their relation to the CRF used. With the use of ResNet, a model for estimating the CRF for each frame of a video can be derived, these per-frame predictions are further used in a video classification model which performs a total CRF prediction for a given video. The results show that it is possible to find a relation between the visual encoding artifacts and CRF used. The top-5 accuracy achieved for the model is at 61.9% with the use of limited training data. Given that today’s parametric bitrate based models for quality have no information about coding complexity, even a rough estimate of the CRF could improve the precision of them.

Microcirculation plays an essential and functional role in the human body and reflects people’s physical status with microscopic detail. For peripheral microcirculation, nail-fold microscopy is a convenient and non-invasive tool since the capillaries in the nail-fold are well arranged and parallel to the skin, which is advantageous for microscopic visualization. Further, nail-fold capillaroscopy information is widely useful. In diagnosis, various diseases such as systemic lupus erythematosus and cardiac diseases can be detected and predicted at an early stage with capillaroscopic patterns and capillary blood velocity. For medical experiments, capillaroscopic information can be used to monitor drug effects and other medical treatments. Though nail-fold capillaroscopy is of significant convenience, it is not widely used. Currently, there is no commercial product with those functions due to the limitations of the equipment, such as microscope resolution and lens magnification. Besides, there is no concrete standard for measurement procedures or objective rules for quantitive data analysis. This thesis proposes a reliable system estimating nail-fold capillary blood flow velocity. It is tested and applied to the microscope from Optilia. In this work, various image and video processing methods are discussed in detail and tested in practice. Taking computational load and equipment limitations into consideration, the system applies frame enhancement and video stabilization. It uses dual-window and correlation methods to estimate the velocity of red blood cells in nail-fold capillaries. In order to test the reliability of the system, the obtained results are compared with the outcome of direct observation. It turns out that the chosen methods employed in the system provide rational results within 5 pixel bias.

Recently, strategies of National Basketball Association (NBA) teams have evolved with the skillsets of players and the emergence of advanced analytics. This has led to a more free-flowing game in which traditional positions and play calls have been replaced with player archetypes and read-and-react offensives that operate off a variety of isolated actions. The introduction of position tracking technology by SportVU has aided the analysis of these patterns by offering a vast dataset of on-court behavior. There have been numerous attempts to identify and classify patterns by evaluating the outcomes of offensive and defensive strategies associated with actions within this dataset, a job currently done manually by reviewing game tape. Some of these classification attempts have used supervised techniques that begin with labeled sets of plays and feature sets to automate the detection of future cases. Increasingly, however, deep learning approaches such as convolutional neural networks have been used in conjunction with player trajectory images generated from positional data. This enables classification to occur in a bottom-up manner, potentially discerning unexpected patterns. Others have shifted focus from classification, instead using this positional data to evaluate the success of a given possession based on spatial factors such as defender proximity and player factors such as role or skillset. While play/action detection, classification and analysis have each been addressed in literature, a comprehensive approach that accounts for modern trends is still lacking. In this paper, we discuss various approaches to action detection and analysis and ultimately propose an outline for a deep learning approach of identification and analysis resulting in a queryable dataset complete with shot evaluations, thus combining multiple contributions into a serviceable tool capable of assisting and automating much of the work currently done by NBA professionals.

Cette thèse porte sur la similitude entre un fluide et une foule et sur l'adaptation de l’algorithme de Particle Video pour le suivi et l'analyse de foule, ce qui aboutit à la conception d'un système complet pour l'analyse de la foule. Cette thèse en étudie trois aspects : la détection de la foule, l'estimation de sa densité et le tracking des flux afin d'obtenir des caractéristiques de comportement.L’algorithme de détection de la foule est une méthode totalement non supervisée pour la détection et la localisation des foules denses dans des images non-contextualisées. Après avoir calculé des vecteurs de features multi-échelles, une classification binaire est effectuée afin d'identifier la foule et l'arrière-plan.L'algorithme d'estimation de densité s'attaque au problème de l'apprentissage de modèles de régression dans le cas de larges foules denses. L'apprentissage est alors impossible sur données réelles car la vérité terrain est indisponible. Notre méthode repose donc sur l'utilisation de données synthétiques pour la phase d'apprentissage et prouve que le modèle de régression obtenu est valable sur données réelles.Pour notre adaptation de l’algorithme de Particle Video nous considérons le nuage de particules comme statistiquement représentatif de la foule. De ce fait, chaque particule possède des propriétés physiques qui nous permettent d'évaluer la validité de son comportement en fonction de celui attendu d'un piéton et d’optimiser son mouvement guidé par le flot optique. Trois applications en découlent : détection des zones d’entrée-sortie de la foule, détection des occlusions dynamiques et mise en relation des zones d'entrée et de sortie selon les flux de piétons. / This thesis focuses on the similarity between a fluid and a crowd and on the adaptation of the particle video algorithm for crowd tracking and analysis. This interrogation ended up with the design of a complete system for crowd analysis out of which, this thesis has addressed three main problems: the detection of the crowd, the estimation of its density and the tracking of the flow in order to derive some behavior features.The contribution to crowd detection introduces a totally unsupervised method for the detection and location of dense crowds in images without context-awareness. After retrieving multi-scale texture-related feature vectors from the image, a binary classification is conducted to identify the crowd and the background.The density estimation algorithm is tackling the problem of learning regression models when it comes to large dense crowds. In such cases, the learning is impossible on real data as the ground truth is not available. Our method relies on the use of synthetic data for the learning phase and proves that the regression model obtained is valid for a use on real data.Our adaptation of the particle video algorithm leads us to consider the cloud of particles as statistically representative of the crowd. Therefore, each particle has physical properties that enable us to assess the validity of its behavior according to the one expected from a pedestrian, and to optimize its motion guided by the optical flow. This leads us to three applications: the detection of the entry and exit areas of the crowd in the image, the detection of dynamic occlusions and the possibility to link entry areas with exit ones, according to the flow of the pedestrians.

La thèse donne à voir et à comprendre en quoi consistent les organisations et l’expérience du travail dans les organisations à partir de l’analyse d’interactions verbales, corporelles et matérielles filmées dans des bureaux. Développant une approche praxéologique originale du côté de la recherche sur les organisations, nous contribuons en outre aux travaux sur les interactions en interrogeant la présence de l’institution dans des formats interactionnels, dans une démarche comparative que permet le corpus. L’exposition de notre ancrage théorique au croisement de ces courants nous conduit à une question épistémologique : est-il possible d’extraire des interactions en coprésence le Quoi du travail d’organisation, tel que l’ethnométhodologie a pu le formuler pour le travail professionnel ? Les résultats empiriques de l’enquête sont ensuite présentés dans cinq chapitres, chacun consacré à un phénomène ou moment de la vie dans les bureaux : les ouvertures des visites, les clôtures des visites, les appels téléphoniques pendant une interaction en coprésence, la mobilisation dans l’interaction du dispositif vidéo, et enfin les réajustements du cadre de participation. La comparaison des différents environnements de travail, des régularités au sein de chacun et entre eux, permet certaines découvertes. / The thesis endeavours to show and understand the very stuff of organisations and the experience of work in organisations, starting from the analysis of verbal, embodied and material interactions filmed in offices. Developing a praxeological, original approach within theories of organisation, we also aim to contribute to research on interactions by putting to question the relevance of institution within interactional patterns, through the comparative approach enabled by our corpus. A theoretical anchorage at the crossroads of these currents being set, we are lead to an epistemological question: is it possible to extract out of copresent interactions the What of organizing work, the way ethnomethodology did with studies of work? We present in the next five chapters our empirical results, each concerning one phenomenon or sequence of work in offices: opening a visit, closing a visit, answering an incoming phonecall during a copresent interaction, formulating the video cameras in interaction, and finally reajusting participation frame. Comparison of various work settings, of regularities between and within them, leads to some discoveries.

京都大学 / 0048 / 新制・課程博士 / 博士(理学) / 甲第20048号 / 理博第4233号 / 新制||理||1609(附属図書館) / 京都大学大学院理学研究科生物科学専攻 / (主査)教授 中川 尚史, 教授 沼田 英治, 教授 中務 真人 / 学位規則第4条第1項該当 / Doctor of Science / Kyoto University / DFAM

A variety of tools and techniques have been developed to measure the movements of the vocal tract, specifically of the tongue and lips. In recent years, computer technology has allowed for extensive exploration of these precise movements and for the development of speech recognition systems. However, there has been relatively little work on the combination of visible facial movements and internal articulatory activity. In this study, two different technologies were used to explore the internal and external movements of speech production in eight speakers: palatometry quantified tongue contact patterns and computerized video image analysis was used to derive lip shape parameters. Results showed that the lip measures used here cannot predict the identity of phonemes in all speakers as well as the tongue contact patterns can. Results also indicated that the data from lip measures were strongly influenced by who the speaker was, whereas the palatometric data were not. These results suggest that more variation exists in lip shape than in tongue contact patterns during speech production. Understanding more about lip measures and vocal tract movement during speech production may potentially benefit the area of speechreading; however, more research is needed to refine the procedures used.

Teacher education programs that aspire to helping pre-service teachers develop expertise must help students engage in deliberate practice along dimensions of teaching expertise. However, field teaching experiences often lack the quantity and quality of feedback that is needed to help students engage in meaningful teaching practice. The limited availability of supervising teachers makes it difficult to personally observe and evaluate each student teacher's field teaching performances. Furthermore, when a supervising teacher debriefs such an observation, the supervising teacher and student may struggle to communicate meaningfully about the teaching performance. This is because the student teacher and supervisor often have very different perceptions of the same teaching performance. Video analysis tools show promise for improving the quality of feedback student teachers receive in their teaching performance by providing a common reference for evaluative debriefing and allowing students to generate their own feedback by coding videos of their own teaching. This study investigates the reliability of pre-service teacher coding using a video analysis tool. This study found that students were moderately reliable coders when coding video of an expert teacher (49%-68%). However, when the reliability of student coding of their own teaching videos was audited, students showed a high degree of accuracy (91%). These contrasting findings suggest that coding reliability scores may not be simple indicators of student understanding of the teaching competencies represented by a coding scheme. Instead, reliability scores may also be subject to the influence of extraneous factors. For example, reliability scores in this study were influenced by differences in the technical aspects of how students implemented the coding system. Furthermore, reliability scores were influenced by how coding proficiency was measured. Because this study also suggests that students can be taught to improve their coding reliability, further research may improve reliability scores"-and make them a more valid reflection of student understanding of teaching competency-"by training students about the technical aspects of implementing a coding system.

Research states that reflection is the foundation for improved teaching (Dewey, 1933). As a result, educators have used many methods to facilitate teacher reflections. Some of these methods include keeping reflective journals, conducting peer teaching sessions, providing written feedback, giving lesson critiques, conducting action research projects, and using reflective conferences (Cook, Young & Evenson, 2001). As video has become more accessible, educators have also become interested in using video analysis tools to facilitate teacher reflections. However, very little has been published on how the use of video analysis tools influences teacher reflections. If reflection is the foundation for improved teaching, it is important for educators and researchers to understand how the use of these tools impacts teacher reflections. Therefore, the focus of this study was to understand the experience of a supervisor and student teacher as they used a video analysis tool to reflect on teaching. The researcher included thick descriptions of participants' experiences, so researchers and educators interested in using video analysis tools to facilitate reflection will be able to transfer the findings to their individual circumstances. This study compared a student teacher's experience reflecting with a video analysis tool to her experience participating in her department's traditional reflection method, which was a post-lesson conference with her supervisor. The researcher investigated how these reflection methods influenced the student teacher's ability to collect data about her teaching, make judgments about her teaching, design intervention plans for future teaching situations, and evaluate her intervention plans. The participants indicated that both video analysis and the traditional reflection method were beneficial for reflection. Although both methods were beneficial, the student teacher felt that using video analysis to reflect was more useful than the department's traditional reflection method for helping her understand the changes she wanted to make in her teaching. The student teacher felt that video analysis was more useful than the traditional reflection method because it allowed her to notice things that she did not remember or attend to during her lesson, it helped her focus her reflections on specific aspects of her teaching, and the video clips provided evidence to support her discussions with her supervisor.