This thesis deals with designing an automatic generator of deep neural networks for image classification. Theoretical part clarifies what a neural network and formal neuron are. Furthermore, the types of neural network architectures are presented. The focus of this thesis is convolutional neural networks, several pieces of research from this field are mentioned. The practical part of this thesis describes information with regards to the implementation of neural network generator, possible frameworks and programming languages for such implementation. Brief description of the implementation itself is presented as well as implemented layers. Generated neural networks are tested on Google-Landmarks dataset and results are commented upon.

Tato práce se zaměřuje na problém počítání vozidel v statickém obraze bez znalosti geometrických vlastností scény. V rámci řešení bylo implementováno a natrénováno 5 architektur konvolučních neuronových sítí. Také byl pořízen rozsáhlý dataset s 19 310 snímky pořízených z 12pohledů a zachycujících 7 různých scén. Použité konvoluční sítě mapují vstupní vzorek na mapu hustoty vozidel, ze které lze získat jejich počet a lokalizaci v kontextu vstupního snímku. Hlavním přínosem této práce je porovnání a aplikace dosavadních nejlepších řešení pro počítání objektů v obraze. Většina z těchto architektur byla navržena pro počítání lidí v obraze, proto musely být uzpůsobeny pro potřeby počítání vozidel v statickém obraze. Natrénované modely jsou vyhodnoceny GAME metrikou na TRANCOS datasetu a na velkém spojeném datasetu. Dosažené výsledky všech modelů jsou následně popsány a porovnány.

Si de nombreux travaux ont été consacrés aux représentations spatiales chez le jeune adulte, la nature des modèles spatiaux, les processus qui président à leur construction et la façon dont ils se développent sont encore loin d'être compris. L'originalité de cette thèse tient au fait d'étudier conjointement les composantes cognitives et langagières dans l'acquisition des représentations d'itinéraires virtuels par des enfants (de 5 à 11 ans) et des adultes, ainsi que les différences individuelles liées à des capacités générales variées. Dans une première partie, la thèse présente les principaux concepts de la cognition spatiale issus des travaux menés chez l'adulte ainsi que l'état des connaissances théoriques et empiriques actuelles sur le développement des représentations spatiales chez l'enfant. Un chapitre s'intéresse ensuite au rôle du langage dans la construction des représentations spatiales et un autre à celui de la mémoire de travail. Afin de mieux comprendre le type de représentation qu'un enfant élabore au cours de son développement, une deuxième partie de la thèse présente trois expériences étudiant le développement des connaissances sur les repères et la route. Les deux premières études ont permis d'observer une augmentation qualitative et quantitative de la connaissance des repères, c'est-à-dire des entités spécifiques qui jalonnent un itinéraire, mais également de la connaissance de la succession de ces repères et des directions empruntées. Le rôle particulier des repères situés à un changement de direction est attesté chez l'adulte comme chez l'enfant. L'augmentation de ces connaissances avec l'âge est observée avec des tâches de production et de reconnaissance, aussi bien verbales que non-verbales. Ces résultats suggèrent l'existence d'une seule représentation commune ou de deux formats de représentations fortement reliés. Le lien important entre les informations verbales et non-verbales dans les représentations est attesté par l'observation d'un biais de type sémantique dans la reconnaissance visuelle de repères. Cependant, l'analyse des différences interindividuelles a mis en évidence le rôle de capacités visuo-spatiales telles que la perception des directions, mais pas d'influence des capacités langagières sur la capacité de représentation d'itinéraire. Une troisième étude explore le rôle des composantes verbales et visuo-spatiales de la mémoire de travail dans le développement des représentations spatiales au moyen d'un paradigme de double tâche lors de la mémorisation d'itinéraires. L'implication, notamment de la composante spatiale de la mémoire de travail au cours de la mémorisation d'itinéraire, est mise en évidence chez l'enfant. Ce résultat renforce l'idée de la dominance d'une forme de codage visuo-spatial dans le développement de la représentation qui évoluerait au profit de codages plus verbaux ou mixtes. En conclusion, cette thèse montre le développement de la capacité à se représenter un itinéraire au cours de l'enfance, attesté par des tâches de nature et de format variés. Si cette représentation semble impliquer à la fois des composantes verbales et non-verbales, ces dernières semblent être plus importantes chez l'enfant. La dernière partie de la thèse propose une discussion des implications de ces résultats pour le développement de la cognition spatiale chez l'enfant, ainsi que des perspectives pour les recherches futures. / Although many studies have investigated spatial representation in young adults, little is still known about the processes underlying how they construct spatial models, the nature of these models, and how they develop in children. The originality of this thesis is two-fold: it studies both cognitive and linguistic processes involved in how children (5 to 11 years) and adults construct representation of virtual routes; it also examines individual differences in these processes. The first part of this thesis begins with a chapter that presents the main concepts underlying spatial cognition, as well as some experimental evidence concerning adults' spatial knowledge and the development of this knowledge during childhood. A second chapter then focuses on the role of language and a third one on the role of working memory in the construction of spatial representations. In order to understand how children construct spatial representation during development, a second part of the thesis presents three experiments investigating the development of landmark and route knowledge. The first two studies show developments in the quality and quantity of knowledge concerning both landmarks (i.e. specific entities encountered along the route) and the route (i.e. the sequential order of actions and landmarks). They also provide evidence supporting the specific role of landmarks associated with changes of direction ("decisional" landmarks) in children and adults. Developmental changes in spatial knowledge were assessed by both verbal and non-verbal measures, suggesting the existence of a unique representation or of two representations that are strongly related. The relation between verbal and non-verbal information in participants' representations is evidenced by their bias toward choosing a related landmark of the same semantic category, regardless of its visual characteristics. Nevertheless, analyses show that visuo-spatial abilities such as the perception of directions, but not verbal abilities, play a main role in accounting for individual differences. The third study, investigates verbal and visuo-spatial components of working memory, using a dual task paradigm in which participants performed a verbal or spatial interference task while memorizing routes. The results support the idea that representing itineraries mostly involves a spatial mode of encoding in children and a more verbal or mixed encoding in adults. To conclude, this thesis shows a development in children's capacity to build spatial representations of virtual routes. Although their representation seems to integrate both verbal and non-verbal components, non-verbal abilities appear to be most essential for children. The last part of the thesis discusses the implications of our results for our understanding of the development of spatial cognition in children, as well as future perspectives and conclusions.

Introduction : L'objectif de cette étude est de comparer la différence entre l'analyse céphalométrique manuelle et l'analyse automatisée par l’intelligence artificielle afin de confirmer la fiabilité de cette dernière. Notre hypothèse de recherche est que la technique manuelle est la plus fiable des deux méthodes. Méthode : Un total de 99 radiographies céphalométriques latérales sont recueillies. Des tracés par technique manuelle (MT) et par localisation automatisée par intelligence artificielle (AI) sont réalisés pour toutes les radiographies. La localisation de 29 points céphalométriques couramment utilisés est comparée entre les deux groupes. L'erreur radiale moyenne (MRE) et un taux de détection réussie (SDR) de 2 mm sont utilisés pour comparer les deux groupes. Le logiciel AudaxCeph version 6.2.57.4225 est utilisé pour l'analyse manuelle et l'analyse AI. Résultats : Le MRE et SDR pour le test de fiabilité inter-examinateur sont respectivement de 0,87 ± 0,61mm et 95%. Pour la comparaison entre la technique manuelle MT et le repérage par intelligence artificielle AI, le MRE et SDR pour tous les repères sont respectivement de 1,48 ± 1,42 mm et 78 %. Lorsque les repères dentaires sont exclus, le MRE diminue à 1,33 ± 1,39 mm et le SDR augmente à 84 %. Lorsque seuls les repères des tissus durs sont inclus (excluant les points des tissus mous et dentaires), le MRE diminue encore à 1,25 ± 1,09 mm et le SDR augmente à 85 %. Lorsque seuls les points de repère des tissus mous sont inclus, le MRE augmente à 1,68 ± 1,89 mm et le SDR diminue à 78 %. Conclusion: La performance du logiciel est similaire à celles précédemment rapportée dans la littérature pour des logiciels utilisant un cadre de modélisation similaire. Nos résultats révèlent que le repérage manuel a donné lieu à une plus grande précision. Le logiciel a obtenu de très bons résultats pour les points de tissus durs, mais sa précision a diminué pour les tissus mous et dentaires. Nous concluons que cette technologie est très prometteuse pour une application en milieu clinique sous la supervision du docteur. / Introduction: The objective of this study is to compare the difference between manual cephalometric analysis and automatic analysis by artificial intelligence to confirm the reliability of the latter. Our research hypothesis is that the manual technique is the most reliable of the methods and is still considered the gold standard. Method: A total of 99 lateral cephalometric radiographs were collected in this study. Manual technique (MT) and automatic localization by artificial intelligence (AI) tracings were performed for all radiographs. The localization of 29 commonly used landmarks were compared between both groups. Mean radial error (MRE) and a successful detection rate (SDR) of 2mm were used to compare both groups. AudaxCeph software version 6.2.57.4225 (Audax d.o.o., Ljubljana, Slovenia) was used for both manual and AI analysis. Results: The MRE and SDR for the inter-examinator reliability test were 0.87 ± 0.61mm and 95% respectively. For the comparison between the manual technique MT and landmarking with artificial intelligence AI, the MRE and SDR for all landmarks were 1.48 ± 1.42mm and 78% respectively. When dental landmarks are excluded, the MRE decreases to 1.33 ± 1.39mm and the SDR increases to 84%. When only hard tissue landmarks are included (excluding soft tissue and dental points) the MRE decreases further to 1.25 ± 1.09mm and the SDR increases to 85%. When only soft tissue landmarks are included the MRE increases to 1.68 ± 1.89mm and the SDR decreases to 78%. Conclusion: The software performed similarly to what was previously reported in literature for software that use analogous modeling framework. Comparing the software’s landmarking to manual landmarking our results reveal that the manual landmarking resulted in higher accuracy. The software operated very well for hard tissue points, but its accuracy went down for soft and dental tissue. Our conclusion is this technology shows great promise for application in clinical settings under the doctor’s supervision.

Superpixels are the results of an image oversegmentation. They are an established intermediate level image representation and used for various applications including object detection, 3d reconstruction and semantic segmentation. While there are various approaches to create such segmentations, there is a lack of knowledge about their properties. In particular, there are contradicting results published in the literature. This thesis identifies segmentation quality, stability, compactness and runtime to be important properties of superpixel segmentation algorithms. While for some of these properties there are established evaluation methodologies available, this is not the case for segmentation stability and compactness. Therefore, this thesis presents two novel metrics for their evaluation based on ground truth optical flow. These two metrics are used together with other novel and existing measures to create a standardized benchmark for superpixel algorithms. This benchmark is used for extensive comparison of available algorithms. The evaluation results motivate two novel segmentation algorithms that better balance trade-offs of existing algorithms: The proposed Preemptive SLIC algorithm incorporates a local preemption criterion in the established SLIC algorithm and saves about 80 % of the runtime. The proposed Compact Watershed algorithm combines Seeded Watershed segmentation with compactness constraints to create regularly shaped, compact superpixels at the even higher speed of the plain watershed transformation. Operating autonomous systems over the course of days, weeks or months, based on visual navigation, requires repeated recognition of places despite severe appearance changes as they are for example induced by illumination changes, day-night cycles, changing weather or seasons - a severe problem for existing methods. Therefore, the second part of this thesis presents two novel approaches that incorporate superpixel segmentations in place recognition in changing environments. The first novel approach is the learning of systematic appearance changes. Instead of matching images between, for example, summer and winter directly, an additional prediction step is proposed. Based on superpixel vocabularies, a predicted image is generated that shows, how the summer scene could look like in winter or vice versa. The presented results show that, if certain assumptions on the appearance changes and the available training data are met, existing holistic place recognition approaches can benefit from this additional prediction step. Holistic approaches to place recognition are known to fail in presence of viewpoint changes. Therefore, this thesis presents a new place recognition system based on local landmarks and Star-Hough. Star-Hough is a novel approach to incorporate the spatial arrangement of local image features in the computation of image similarities. It is based on star graph models and Hough voting and particularly suited for local features with low spatial precision and high outlier rates as they are expected in the presence of appearance changes. The novel landmarks are a combination of local region detectors and descriptors based on convolutional neural networks. This thesis presents and evaluates several new approaches to incorporate superpixel segmentations in local region detection. While the proposed system can be used with different types of local regions, in particular the combination with regions obtained from the novel multiscale superpixel grid shows to perform superior to the state of the art methods - a promising basis for practical applications.

Is it important that a player can navigate easily through a level? To get the answer ten participants were tested during a short play-session of a game made for the study. By looking at prior research, themes could be found that were all common. To learn the player to recognize what a goal looks like, to get the player to understand what the goal is. That the player can navigate to the goal and that the level should have flow is also important. This study shows that what earlier research suggests seems to have merit. Of all participants that played the game, almost everyone that played where navigational techniques were implemented spoke of how they felt certain in where to go.

Superpixels are the results of an image oversegmentation. They are an established intermediate level image representation and used for various applications including object detection, 3d reconstruction and semantic segmentation. While there are various approaches to create such segmentations, there is a lack of knowledge about their properties. In particular, there are contradicting results published in the literature. This thesis identifies segmentation quality, stability, compactness and runtime to be important properties of superpixel segmentation algorithms. While for some of these properties there are established evaluation methodologies available, this is not the case for segmentation stability and compactness. Therefore, this thesis presents two novel metrics for their evaluation based on ground truth optical flow. These two metrics are used together with other novel and existing measures to create a standardized benchmark for superpixel algorithms. This benchmark is used for extensive comparison of available algorithms. The evaluation results motivate two novel segmentation algorithms that better balance trade-offs of existing algorithms: The proposed Preemptive SLIC algorithm incorporates a local preemption criterion in the established SLIC algorithm and saves about 80 % of the runtime. The proposed Compact Watershed algorithm combines Seeded Watershed segmentation with compactness constraints to create regularly shaped, compact superpixels at the even higher speed of the plain watershed transformation. Operating autonomous systems over the course of days, weeks or months, based on visual navigation, requires repeated recognition of places despite severe appearance changes as they are for example induced by illumination changes, day-night cycles, changing weather or seasons - a severe problem for existing methods. Therefore, the second part of this thesis presents two novel approaches that incorporate superpixel segmentations in place recognition in changing environments. The first novel approach is the learning of systematic appearance changes. Instead of matching images between, for example, summer and winter directly, an additional prediction step is proposed. Based on superpixel vocabularies, a predicted image is generated that shows, how the summer scene could look like in winter or vice versa. The presented results show that, if certain assumptions on the appearance changes and the available training data are met, existing holistic place recognition approaches can benefit from this additional prediction step. Holistic approaches to place recognition are known to fail in presence of viewpoint changes. Therefore, this thesis presents a new place recognition system based on local landmarks and Star-Hough. Star-Hough is a novel approach to incorporate the spatial arrangement of local image features in the computation of image similarities. It is based on star graph models and Hough voting and particularly suited for local features with low spatial precision and high outlier rates as they are expected in the presence of appearance changes. The novel landmarks are a combination of local region detectors and descriptors based on convolutional neural networks. This thesis presents and evaluates several new approaches to incorporate superpixel segmentations in local region detection. While the proposed system can be used with different types of local regions, in particular the combination with regions obtained from the novel multiscale superpixel grid shows to perform superior to the state of the art methods - a promising basis for practical applications.

Human pose estimation (HPE) constitutes a fundamental problem within the domain of computer vision, finding applications in diverse fields like motion analysis and human-computer interaction. This paper introduces innovative methodologies aimed at enhancing the accuracy and robustness of 3D joint estimation. Through the integration of Variational Autoencoders (VAEs), pertinent information is extracted from depth maps, even in the presence of inevitable image-capturing inconsistencies. This concept is enhanced through the introduction of noise to the body or specific regions surrounding key joints. The deliberate introduction of noise to these areas enables the VAE to acquire a robust representation that captures authentic pose-related patterns. Moreover, the introduction of a localised mask as a constraint in the loss function ensures the model predominantly relies on pose-related cues while disregarding potential confounding factors that may hinder the compact representation of accurate human pose information. Delving into the latent space modulation further, a novel model architecture is devised, joining a VAE and fully connected network into a multi-task joint training objective. In this framework, the VAE and regressor harmoniously influence the latent representations for accurate joint detection and localisation. By combining the multi-task model with the loss function constraint, this study attains results that compete with state-of-the-art techniques. These findings underscore the significance of leveraging latent space modulation and customised loss functions to address challenging human poses. Additionally, these novel methodologies pave the way for future explorations and provide prospects for advancing HPE. Subsequent research endeavours may optimising these techniques, evaluating their performance across diverse datasets, and exploring potential extensions to unravel further insights and advancements in the field. / Human pose estimation (HPE) är ett grundläggande problem inom datorseende och används inom områden som rörelseanalys och människa-datorinteraktion. I detta arbete introduceras innovativa metoder som syftar till att förbättra noggrannheten och robustheten i 3D-leduppskattning. Genom att integrera variationsautokodare (eng. variational autoencoder, VAE) extraheras relevant information från djupkartor, trots närvaro av inkonsekventa avvikelser i bilden. Dessa avvikelser förstärks genom att applicera brus på kroppen eller på specifika regioner som omger viktiga leder. Det avsiktliga införandet av brus i dessa områden gör det möjligt för VAE att lära sig en robust representation som fångar autentiska poseringsrelaterade mönster. Dessutom införs en lokaliserad mask som en begränsning i förlustfunktionen, vilket säkerställer att modellen främst förlitar sig på poseringsrelaterade signaler samtidigt som potentiella störande faktorer som hindrar den kompakta representationen av korrekt mänsklig poseringsinformation bortses ifrån. Genom att fördjupa sig ytterligare i den latenta rumsmoduleringen har en ny modellarkitektur tagits fram som förenar en VAE och ett fullständigt anslutet nätverk i en fleruppgiftsmodell. I detta ramverk påverkar VAE och det fullständigt ansluta nätverket de latenta representationerna på ett harmoniskt sätt för att uppnå korrekt leddetektering och lokalisering. Genom att kombinera fleruppgiftsmodellen med förlustfunktionsbegränsningen uppnår denna studie resultat som konkurrerar med toppmoderna tekniker. Dessa resultat understryker betydelsen av att utnyttja latent rymdmodulering och anpassade förlustfunktioner för att hantera utmanande mänskliga poser. Dessutom banar dessa nya metoder väg för framtida utveckling inom uppskattning av HPE. Efterföljande forskningsinsatser kan optimera dessa tekniker, utvärdera deras prestanda över olika datamängder och utforska potentiella tillägg för att avslöja ytterligare insikter och framsteg inom området.

No Abstract Provided / Thesis / Bachelor of Arts (BA)