Analysis and extraction of features in video streaming refer to the process of identifying and extracting specific characteristics or patterns from a video stream that can be used for various purposes, such as object detection, recognition, and tracking, as well as video compression, indexing, and retrieval. The extracted features can be used for different purposes, depending on the specific needs and requirements of the application. End-to-end streaming latency refers to the delay between the time a video frame is captured and the time it is displayed on the user’s device. Analysis and extraction of features in video streaming can be used to measure end-to-end streaming latency by extracting specific characteristics or patterns from the video stream that indicate the start and end points of the video stream and the time stamps of each frame. In this work we propose a simple but effective way to measure the end-to-end streaming latency by using object detection and image-to-text conversion, both tasks based on the extraction of features of the underlying content. Shot boundary detection is the process of identifying the boundaries between shots in a video stream. Shot boundary detection is an important task in video processing, as it is used for various applications, such as video editing, indexing, retrieval, and summarization. Analysis and extraction of features in video streaming can be used for shot boundary detection by extracting specific characteristics or patterns from the video stream that indicate changes in the visual and audio content. Once these features are extracted, various techniques can be used to detect shot boundaries, such as thresholding, clustering, and machine learning algorithms. In this work, we analyze state-of-the-art deep learning algorithms for shot boundary detection tasks and datasets and propose several new models that improve the efficiency of the last state-of-the-art models meanwhile keeping or even improving the resulting metrics. Video temporal segmentation in scenes is the process of dividing a video stream into coherent temporal segments grouping all shots that are visually and semantically related to each other. In this work, we take advantage of the improvements done in the task of shot boundary detection to propose a foundational model in the task of segmenting the video into scenes, from a previous segmentation in shots. We propose a model based on visual similarity and we also contribute with a specific dataset for the task.

Discovering the crucial factors that contribute to goal success in sports analytics, this thesis aimsto utilize Random Forest classification to predict the outcome of shots and pre-shot events in powerplay situations. Through three experiments, the study evaluated the use of shots, shots with pre-shotevents, and shots with pre-shot events over sections. The first experiment used only shots, while thesecond experiment focused on shots with pre-shot events, where both compared it with shots over anexpected goal value of 0.08 or higher. The third experiment examined shots with pre-shot events acrossdifferent sections. Our findings demonstrated that the models in our experiments achieved accuracyscores ranging from 78% to 96% and F1 scores between 0% and 24%. Notably, the models in experiment3 demonstrated lower recall scores. The feature importance analysis revealed that pre-shotevents played a significant role in the predictive models of the second and third experiments, indicatingtheir substantial impact on the outcomes. A noteworthy conclusion arising from the discussion isthe recommendation for future research to conduct a more comprehensive exploration into the impactof pre-shot events, given their demonstrated significance in predicting goals. Such an investigation isdeemed necessary and justified.

Sammanfattning Väldigt lite är känt om hur olika träningsmetoder påverkar teknisk utveckling hos innebandyspelare. Syftet med studien var att undersöka hur olika typer av smålagsspel i innebandy påverkar spelvolymen samt antal avslut och avslut på mål. Frågeställningarna för arbetet var: skiljer sig spelvolymen i innebandyns smålagsspel mellan spel två mot två (2v2), 3v3, 4v4 och 5v5? Skiljer sig antalet avslut i innebandyns smålagsspel mellan spel 2v2, 3v3, 4v4 och 5v5? Skiljer sig antalet avslut på mål i innebandyns smålagsspel mellan spel 2v2, 3v3, 4v4 och 5v5? Metod Två testgrupper rekryterades till studien där grupperna bestod av två olika amatörlag. Det ena laget var ett damveteranlag där 15 spelare deltog, med en medelålder på 38 år, med 2-15 års erfarenhet av organiserad innebandy. Det andra laget var ett pojklag med 21 deltagande spelare, i åldrarna 12-13 år, med 5-6 års erfarenhet av innebandy. Datainsamlingen skedde sedan under respektive lags ordinarie träning där fyra olika spelformer filmades. Ena laget filmades tio gånger för vardera spelform och det andra laget filmades tio gånger för vardera spelform förutom en, där en spelsekvens försvann på grund av oförutsedda händelser. Varje spelsekvens var en minut lång. Därefter analyserades filmmaterialet varpå spelvolym, antal avslut samt avslut på mål registrerades för varje minut spelad i de olika spelformerna. Rådata sammanställdes och genomgick sedan en statistisk analys för att undersöka om spelen skiljde sig signifikant i någon del. Resultat Grupp A visade på signifikant skillnad i spelvolym mellan alla spelformer där den högsta volymen hittades i spel 2v2, därefter följde 3v3, 4v4 och sist 5v5. Alla smålagsspel för grupp A visade signifikant fler avslut och avslut på mål jämfört med spel 5v5 men ingen signifikant skillnad hittades smålagsspelen emellan. I grupp B hittades samma mönster för spelvolym som i grupp A, dock visades ingen signifikant skillnad mellan spelformerna 4v4 och 5v5. Antalet avslut var signifikant högre i alla smålagsspel jämfört med 5v5, dock hittades endast signifikans mellan 2v2 och 5v5 när det kom till avslut på mål. Slutsats Studien indikerar att en minskning i antalet spelare verkar vara ett effektivt sätt att öka frekvensen på de tekniska aktionerna under spel. För teknikträning i spel verkar därför smålagsspel vara att föredra framför fullstort spel 5v5. Mer forskning på området behövs för att öka förståelsen kring möjliga användningsområden för innebandyns smålagsspel. / Aim Very little is known about how different types of training influence technical development in floorball. The purpose of this study was to investigate how volume of play, shots, and shots on goal, were affected in different types of small-sided games (SSG). The research questions this paper sought to answer were: in floorball, does the volume of play differ between the game formats: two versus two (2v2), 3v3, 4v4, and 5v5? In floorball, does the number of shots differ between the game formats: 2v2, 3v3, 4v4, and 5v5? In floorball, does the number of shots on goal differ between the game formats: 2v2, 3v3, 4v4, and 5v5? Method Two amateur teams were recruited for the study. One team was a women’s team of 15 players participating, with an average age of 38, with 2-15 years’ experience of organized floorballpractice. The other team was a boys’ team of 21 players participating, aged 12-13 and 5-6 years’experience of organized floorball practice. The data were collected during one of each team’sregular floorball practice sessions, where video captured the four different SSGs. One team was filmed ten times for one minute for each SSG and the other team was filmed ten times in three SSGs and nine times in one. The videos were analysed for volume of play, as well as registering shots and shots on goal. The compiled data was then statistically analysed in order to see if there were any significant difference between the game formats. Results Group A showed a significant difference in volume of play between all SSGs, where the highest volume of play was found in 2v2, and the rest with a lower volume of play in ascending order. As for shots and shots on goal, group A showed significantly more shots and shots on goal in 2v2, 3v3, and 4v4 compared to 5v5, but no significance was found between the SSGs. Group B showed a similar pattern in volume of play as found in group A, but no significant difference between 4v4 and 5v5. The amount of shots were significantly higher in 2v2, 3v3, and 4v4 compared to 5v5. However, for shots on goal group B only showed 2v2 producing significantly more shots than 5v5. Conclusion The study indicates that reducing the number ofplayers is aneffective wayto increase technical actions. Therefore, SSGs seems to be better for skill acquisition in games than the large game of 5v5. However, more research is needed in this field for further understanding of the uses of SSG:s in floorball.

Title: Intraindividual stability of performance and kinestetic abilities level in wedge play performed on golf players with different performance level Goal: Goal of this thesis is detect an intraindividual stability of full swing performance in wedge play and find out kinestetic abilities level of golf players with different performance level when they control distance of ball flight, club head speed and ball speed. Methods: There were 15 golfers participating in this study (n = 15). Tested subjects are characteristic by these values: age 18,36 ± 2,61, body height 180,86 ± 7,38, body weight 73,21 ± 10,25, HCP HCP -0,47 ± 1,53. Intraindividual stability of performance and kinestetic abilities level was tested by instrument TrackMan. Gained values were processed by statistical methods in Excel. Pearson coeficient was used to find out the dependence between data. Results: We found high level of intraindividual stability from test results. Significant difference was found between in distance, club head speed and ball speed parameters with and without feedback. Significant connetion between intaindividual stability and performance parameters was found only in 1/3 of cases, same as connection between kinethetic ability and performance parameters. Key words: Short game, golf swing, approach shots, abilities

SyfteSyftet med studien var att undersöka komplexitetsnivåns (stillastående handledsskott och handledsskott i rörelse) och träffpunktsplaceringens inverkan på skottsprecisionen hos damelitinnebandyspelare på seniornivå. Vidare syftade studien till att undersöka om det fanns någon inlärningseffekt vid upprepade skott mot samma träffpunkt.MetodÅtta testdeltagare deltog där två test med olika komplexitetsnivåer utfördes. Första testet utfördes som ett stillastående handledsskott (SH) medan det andra testet utfördes som handledsskott i rörelse (HR). Varje deltagare utförde 40 skott per test fördelat på fyra träffpunkter (Tp). Dessa befann sig i målets övre och nedre del. Testerna filmades och analyserades där skottens x- och y-koordinat fastställdes. Testvärdena analyserades sedan i SPSS där Paired-Sample t-test användes samt Pearsons korrelationsanalys.ResultatDet fanns en signifikant skillnad i skottprecision avseende avstånd från träffpunkterna mellan SH och HR (t = -6,68, p = 0,0068). Det förekom en signifikant skillnad i skottprecision vid HR mellan den närmaste och den bortre övre Tp (t = 3,58, p = 0,0090), övriga Tp i närmaste och bortre hörn i de vardera testen uppvisade inga signifikanta skillnader (alla p > 0,05). Inget signifikant samband fanns avseende inlärningseffekt vid SH eller HR (p > 0,05).SlutsatserSkottprecisionen påverkades av komplexitetsnivåerna som fanns mellan SH och HR. Resultatet kan användas av spelare och tränare då spelarna kan utveckla skottprecisionen i rörelsemoment och att skjuta i målets närmaste övre del. Det inträffade ingen inlärning genom att utföra 10 skott mot samma Tp vilket gör att spelarna kan behöva fler rörelseupprepningar för att få en kortvarig inlärning. / Purpose The purpose of the study was to investigate the level of complexity in stationary wrist shot and wrist shot in motion at female elite floorball players. Furthermore, the study aimed to investigate the impact of the placement of the target and investigate whether there was any learning effect in repeated shots at the same target.Methods Eight participants took part of the study were two tests was performed with different complexity level. The first test was performed as a stationary wrist shot (SH) while test two was performed as a wrist shot in motion (HR). In total, each participant performed 40 shots per test divided in four hit targets (Tp). These were placed in the upper and lower part of the goal. The tests were recorded and analyzed afterwards where the x- and y-coordinates of the shots were determined. The test values were then analyzed in SPSS where Paired-Sample t-test was used and Pearsons correlation analysis.Results There was a significant difference in shot precision between HR and SH (t = -6,68, p = 0,0068). There was a significant difference in shot precision in HR between the nearest upper target and the far upper target (t = 3,58, p = 0,0090). The other nearest and farther targets in each test showed no significant difference (p > 0,05). There was no significant learning effect in SH or HR (p > 0,05).Conclusions Shot accuracy was affected by the complexity levels that existed between SH and HR. The result can be used by players and coaches as players can develop shot precision in movement moments and to shoot in the nearest upper target. No learning occurred by performing 10 shots at the same target which means that players may need more movement repetitions to achieve a short learning effect.

As a leading figure of “The Six Generation” directors, Jia Zhangke’s films focus on reality of contemporary Chinese society, and record the lives of people who were left behind after the country’s urbanization process. He depicts a lot of characters who struggle with their lives, and he works to explore one common question throughout all of his films: “where do I belong?” Jia Zhangke uses unique filmmaking techniques in order to emphasize the feelings of people losing their sense of home. In this thesis, I am going to analyze his cinematic style from three perspectives: photography, musical scores and metaphors. In each chapter, I will use one film as the main subject of discussion and reference other films to complement my analysis. / Graduate

Hit registration algorithms in First-Person Shooter games define how the server processes gunfire from clients. Network conditions, such as latency, cause a mismatch between the game worlds observed at the client and the server. To improve the experience for clients when authoritative servers are used, the server attempts to reconcile the differing views when performing hit registration through techniques known as lag compensation. This thesis surveys recent hit registration techniques and discusses how they can be implemented and evaluated with the use of a modern game engine. To this end, a lag compensation model based on animation pose rewind is implemented in Unreal Engine 4. Several programming models described in industry and research are used in the implementation, and experiences from further integrating the techniques into a commercial FPS project are also discussed. To reason about the accuracy of the algorithm, client-server discrepancy metrics are defined, as well as a hit rate metric which expresses the worst-case effect on the shooting experience of a player. Through automated tests, these metrics are used to evaluate the hit registration accuracy. The rewind algorithm was found to make the body-part-specific hit registration function well independently of latency. At high latencies, the rewind algorithm is completely necessary to make sure that clients can still aim at where they perceive their targets to be and expect their hits to be registered. Still, inconsistencies in the results remain, with hit rate values sometimes falling below 50%. This is theorized to be due to fundamental networking mechanisms of the game engine which are difficult to control. This presents a counterpoint to the otherwisegained ease of implementation when using Unreal Engine.

The thesis investigates a publicly available dataset which covers morethan three million events in football matches. The aim of the study isto train machine learning models capable of modeling the relationshipbetween a shot event and its outcome. That is, to predict if a footballshot will result in a goal or not. By representing the shot indifferent ways, the aim is to draw conclusion regarding what elementsof a shot allows for a good prediction of its outcome. The shotrepresentation was varied both by including different numbers of eventspreceding the shot and by varying the set of features describing eachevent.The study shows that the performance of the machine learning modelsbenefit from including events preceding the shot. The highestpredictive performance was achieved by a long short-term memory neuralnetwork trained on the shot event and six events preceding the shot.The features which were found to have the largest positive impact onthe shot events were the precision of the event, the position on thefield and how the player was in contact with the ball. The size of thedataset was also evaluated and the results suggest that it issufficiently large for the size of the networks evaluated.

La construction d'algorithmes classifiant des images à grande échelle est devenue une tache essentielle du fait de la difficulté d'effectuer des recherches dans les immenses collections de données visuelles inetiquetées présentes sur Internet. Nous visons à classifier des images en fonction de leur contenu pour simplifier la gestion de telles bases de données. La classification d'images à grande échelle est un problème complèxe, de par l'importance de la taille des ensembles de données, tant en nombre d'images qu'en nombre de classes. Certaines de ces classes sont dites "fine-grained" (sémantiquement proches les unes des autres) et peuvent même ne contenir aucun représentant étiqueté. Dans cette thèse, nous utilisons des représentations état de l'art d'images et nous concentrons sur des méthodes d'apprentissage efficaces. Nos contributions sont (1) un banc d'essai d'algorithmes d'apprentissage pour la classification à grande échelle et (2) un nouvel algorithme basé sur l'incorporation d'étiquettes pour apprendre sur des données peu abondantes. En premier lieu, nous introduisons un banc d'essai d'algorithmes d'apprentissage pour la classification à grande échelle, dans le cadre entièrement supervisé. Il compare plusieurs fonctions objectifs pour apprendre des classifieurs linéaires, tels que "un contre tous", "multiclasse", "ranking", "ranking pondéré moyen" par descente de gradient stochastique. Ce banc d'essai se conclut en un ensemble de recommandations pour la classification à grande échelle. Avec une simple repondération des données, la stratégie "un contre tous" donne des performances meilleures que toutes les autres. Par ailleurs, en apprentissage en ligne, un pas d'apprentissage assez petit s'avère suffisant pour obtenir des résultats au niveau de l'état de l'art. Enfin, l'arrêt anticipé de la descente de gradient stochastique introduit une régularisation qui améliore la vitesse d'entraînement ainsi que la capacité de régularisation. Deuxièmement, face à des milliers de classes, il est parfois difficile de rassembler suffisamment de données d'entraînement pour chacune des classes. En particulier, certaines classes peuvent être entièrement dénuées d'exemples. En conséquence, nous proposons un nouvel algorithme adapté à ce scénario d'apprentissage dit "zero-shot". notre algorithme utilise des données parallèles, comme les attributs, pour incorporer les classes dans un espace euclidien. Nous introduisons par ailleurs une fonction pour mesurer la compatibilité entre image et étiquette. Les paramètres de cette fonction sont appris en utilisant un objectif de type "ranking". Notre algorithme dépasse l'état de l'art pour l'apprentissage "zero-shot", et fait preuve d'une grande flexibilité en permettant d'incorporer d'autres sources d'information parallèle, comme des hiérarchies. Il permet en outre une transition sans heurt du cas "zero-shot" au cas où peu d'exemples sont disponibles.

La construction d'algorithmes classifiant des images à grande échelle est devenue une t^ache essentielle du fait de la difficulté d'effectuer des recherches dans les immenses collections de données visuelles non-etiquetées présentes sur Internet. L'objetif est de classifier des images en fonction de leur contenu pour simplifier la gestion de telles bases de données. La classification d'images à grande échelle est un problème complexe, de par l'importance de la taille des ensembles de données, tant en nombre d'images qu'en nombre de classes. Certaines de ces classes sont dites "fine-grained" (sémantiquement proches les unes des autres) et peuvent même ne contenir aucun représentant étiqueté. Dans cette thèse, nous utilisons des représentations à l'état de l'art d'images et nous concentrons sur des méthodes d'apprentissage efficaces. Nos contributions sont (1) un banc d'essai d'algorithmes d'apprentissage pour la classification à grande échelle et (2) un nouvel algorithme basé sur l'incorporation d'étiquettes pour apprendre sur des données peu abondantes. En premier lieu, nous introduisons un banc d'essai d'algorithmes d'apprentissage pour la classification à grande échelle, dans un cadre entièrement supervisé. Il compare plusieurs fonctions objectifs pour apprendre des classifieurs linéaires, tels que "un contre tous", "multiclasse", "classement", "classement avec pondération" par descente de gradient stochastique. Ce banc d'essai se conclut en un ensemble de recommandations pour la classification à grande échelle. Avec une simple repondération des données, la stratégie "un contre tous" donne des performances meilleures que toutes les autres. Par ailleurs, en apprentissage en ligne, un pas d'apprentissage assez petit s'avère suffisant pour obtenir des résultats au niveau de l'état de l'art. Enfin, l'arrêt prématuré de la descente de gradient stochastique introduit une régularisation qui améliore la vitesse d'entraînement ainsi que la capacité de régularisation. Deuxièmement, face à des milliers de classes, il est parfois difficile de rassembler suffisamment de données d'entraînement pour chacune des classes. En particulier, certaines classes peuvent être entièrement dénuées d'exemples. En conséquence, nous proposons un nouvel algorithme adapté à ce scénario d'apprentissage dit "zero-shot". Notre algorithme utilise des données parallèles, comme les attributs, pour incorporer les classes dans un espace euclidien. Nous introduisons par ailleurs une fonction pour mesurer la compatibilité entre image et étiquette. Les paramètres de cette fonction sont appris en utilisant un objectif de type "ranking". Notre algorithme dépasse l'état de l'art pour l'apprentissage "zero-shot", et fait preuve d'une grande flexibilité en permettant d'incorporer d'autres sources d'information parallèle, comme des hiérarchies. Il permet en outre une transition sans heurt du cas "zero-shot" au cas où peu d'exemples sont disponibles. / Building algorithms that classify images on a large scale is an essential task due to the difficulty in searching massive amount of unlabeled visual data available on the Internet. We aim at classifying images based on their content to simplify the manageability of such large-scale collections. Large-scale image classification is a difficult problem as datasets are large with respect to both the number of images and the number of classes. Some of these classes are fine grained and they may not contain any labeled representatives. In this thesis, we use state-of-the-art image representations and focus on efficient learning methods. Our contributions are (1) a benchmark of learning algorithms for large scale image classification, and (2) a novel learning algorithm based on label embedding for learning with scarce training data. Firstly, we propose a benchmark of learning algorithms for large scale image classification in the fully supervised setting. It compares several objective functions for learning linear classifiers such as one-vs-rest, multiclass, ranking and weighted average ranking using the stochastic gradient descent optimization. The output of this benchmark is a set of recommendations for large-scale learning. We experimentally show that, online learning is well suited for large-scale image classification. With simple data rebalancing, One-vs-Rest performs better than all other methods. Moreover, in online learning, using a small enough step size with respect to the learning rate is sufficient for state-of-the-art performance. Finally, regularization through early stopping results in fast training and a good generalization performance. Secondly, when dealing with thousands of classes, it is difficult to collect sufficient labeled training data for each class. For some classes we might not even have a single training example. We propose a novel algorithm for this zero-shot learning scenario. Our algorithm uses side information, such as attributes to embed classes in a Euclidean space. We also introduce a function to measure the compatibility between an image and a label. The parameters of this function are learned using a ranking objective. Our algorithm outperforms the state-of-the-art for zero-shot learning. It is flexible and can accommodate other sources of side information such as hierarchies. It also allows for a smooth transition from zero-shot to few-shots learning.