A representação de imagem baseada em superpixels tem se tornado indispensável na melhoria da eficiência em sistemas de Visão Computacional. Reconhecimento de objetos, segmentação, estimativa de profundidade e estimativa de modelo corporal são alguns importantes problemas nos quais superpixels podem ser aplicados. Porém, superpixels podem influenciar a qualidade dos resultados do sistema positiva ou negativamente, dependendo de quão bem eles respeitam as fronteiras dos objetos na imagem. Neste trabalho, é proposto um método iterativo para geração de superpixels, conhecido por IFT-SLIC, baseado em sequências de Transformadas Imagem-Floresta, começando com uma grade regular de sementes. Um procedimento de recomputação de pixels sementes é aplicado a cada iteração, gerando superpixels conexos com melhor aderência às bordas dos objetos presentes na imagem. Os superpixels obtidos via IFT-SLIC correspondem, estruturalmente, a árvores de espalhamento enraizadas nessas sementes, que naturalmente definem superpixels como regiões de pixels fortemente conexas. Comparadas ao Agrupamento Iterativo Linear Simples (SLIC), o IFT-SLIC considera os custos dos caminhos mínimos entre pixels e os centros dos agrupamentos, em vez de suas distâncias diretas. Funções de conexidade não monotonicamente incrementais são exploradas em neste método resultando em melhor desempenho. Estudos experimentais indicam resultados de extração de superpixels superiores pelo método proposto em comparação com o SLIC. Também é analisada a efetividade do IFT-SLIC, em termos de medidas de eficiência e acurácia, em uma aplicação de segmentação do céu em fotos de paisagens. Os resultados mostram que o IFT-SLIC é competitivo com os melhores métodos do estado da arte e superior a muitos outros, motivando seu desenvolvimento para diferentes aplicações. / Image representation based on superpixels has become indispensable for improving efficiency in Computer Vision systems. Object recognition, segmentation, depth estimation, and body model estimation are some important problems where superpixels can be applied. However, superpixels can influence the quality of the system results in a positive or negative manner, depending on how well they respect the object boundaries in the image. In this work, we propose an iterative method for superpixels generation, known as IFT-SLIC, which is based on sequences of Image Foresting Transforms, starting with a regular grid for seed sampling. A seed pixel recomputation procedure is applied per each iteration, generating connected superpixels with a better adherence to objects borders present in the image. The superpixels obtained by IFT-SLIC structurally correspond to spanning trees rooted at those seeds, that naturally define superpixels as regions of strongly connected pixels. Compared to Simple Linear Iterative Clustering (SLIC), IFT-SLIC considers minimum path costs between pixel and cluster centers rather than their direct distances. Non-monotonically increasing connectivity functions are explored in our IFT-SLIC approach leading to improved performance. Experimental results indicate better superpixel extraction by the proposed approach in comparation to that of SLIC. We also analyze the effectiveness of IFT-SLIC, according to efficiency, and accuracy on an application -- namely sky segmentation. The results show that IFT-SLIC can be competitive to the best state-of-the-art methods and superior to many others, which motivates it\'s further development for different applications.

A representação de imagem baseada em superpixels tem se tornado indispensável na melhoria da eficiência em sistemas de Visão Computacional. Reconhecimento de objetos, segmentação, estimativa de profundidade e estimativa de modelo corporal são alguns importantes problemas nos quais superpixels podem ser aplicados. Porém, superpixels podem influenciar a qualidade dos resultados do sistema positiva ou negativamente, dependendo de quão bem eles respeitam as fronteiras dos objetos na imagem. Neste trabalho, é proposto um método iterativo para geração de superpixels, conhecido por IFT-SLIC, baseado em sequências de Transformadas Imagem-Floresta, começando com uma grade regular de sementes. Um procedimento de recomputação de pixels sementes é aplicado a cada iteração, gerando superpixels conexos com melhor aderência às bordas dos objetos presentes na imagem. Os superpixels obtidos via IFT-SLIC correspondem, estruturalmente, a árvores de espalhamento enraizadas nessas sementes, que naturalmente definem superpixels como regiões de pixels fortemente conexas. Comparadas ao Agrupamento Iterativo Linear Simples (SLIC), o IFT-SLIC considera os custos dos caminhos mínimos entre pixels e os centros dos agrupamentos, em vez de suas distâncias diretas. Funções de conexidade não monotonicamente incrementais são exploradas em neste método resultando em melhor desempenho. Estudos experimentais indicam resultados de extração de superpixels superiores pelo método proposto em comparação com o SLIC. Também é analisada a efetividade do IFT-SLIC, em termos de medidas de eficiência e acurácia, em uma aplicação de segmentação do céu em fotos de paisagens. Os resultados mostram que o IFT-SLIC é competitivo com os melhores métodos do estado da arte e superior a muitos outros, motivando seu desenvolvimento para diferentes aplicações. / Image representation based on superpixels has become indispensable for improving efficiency in Computer Vision systems. Object recognition, segmentation, depth estimation, and body model estimation are some important problems where superpixels can be applied. However, superpixels can influence the quality of the system results in a positive or negative manner, depending on how well they respect the object boundaries in the image. In this work, we propose an iterative method for superpixels generation, known as IFT-SLIC, which is based on sequences of Image Foresting Transforms, starting with a regular grid for seed sampling. A seed pixel recomputation procedure is applied per each iteration, generating connected superpixels with a better adherence to objects borders present in the image. The superpixels obtained by IFT-SLIC structurally correspond to spanning trees rooted at those seeds, that naturally define superpixels as regions of strongly connected pixels. Compared to Simple Linear Iterative Clustering (SLIC), IFT-SLIC considers minimum path costs between pixel and cluster centers rather than their direct distances. Non-monotonically increasing connectivity functions are explored in our IFT-SLIC approach leading to improved performance. Experimental results indicate better superpixel extraction by the proposed approach in comparation to that of SLIC. We also analyze the effectiveness of IFT-SLIC, according to efficiency, and accuracy on an application -- namely sky segmentation. The results show that IFT-SLIC can be competitive to the best state-of-the-art methods and superior to many others, which motivates it\'s further development for different applications.

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.

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.

In recent years, Polarimetric Synthetic Aperture Radar (PolSAR) has been one of the most important instruments for earth observation, and is increasingly used in various remote sensing applications. Statistical modelling and scattering analysis are two main ways for PolSAR data interpretation, and have been intensively investigated in the past two decades. Moreover, spatial analysis was applied in the analysis of PolSAR data and found to be beneficial to achieve more accurate interpretation results. This thesis focuses on extracting typical spatial information, i.e., edges and regions by exploring the statistical characteristics of PolSAR data. The existing spatial analysing methods are mainly based on the complex Wishart distribution, which well characterizes the inherent statistical features in homogeneous areas. However, the non-Gaussian models can give better representation of the PolSAR statistics, and therefore have the potential to improve the performance of spatial analysis, especially in heterogeneous areas. In addition, the traditional fixed-shape windows cannot accurately estimate the distribution parameter in some complicated areas, leading to the loss of the refined spatial details. Furthermore, many of the existing methods are not spatially adaptive so that the obtained results are promising in some areas whereas unsatisfactory in other areas. Therefore, this thesis is dedicated to extracting spatial information by applying the non-Gaussian statistical models and spatially adaptive strategies. The specific objectives of the thesis include: (1) to develop reliable edge detection method, (2) to develop spatially adaptive superpixel generation method, and (3) to investigate a new framework of region-based segmentation. Automatic edge detection plays a fundamental role in spatial analysis, whereas the performance of classical PolSAR edge detection methods is limited by the fixed-shape windows. Paper 1 investigates an enhanced edge detection method using the proposed directional span-driven adaptive (DSDA) window. The DSDA window has variable sizes and flexible shapes, and can overcome the limitation of fixed-shape windows by adaptively selecting homogeneous samples. The spherically invariant random vector (SIRV) product model is adopted to characterize the PolSAR data, and a span ratio is combined with the SIRV distance to highlight the dissimilarity measure. The experimental results demonstrated that the proposed method can detect not only the obvious edges, but also the tiny and inconspicuous edges in heterogeneous areas. Edge detection and region segmentation are two important aspects of spatial analysis. As to the region segmentation, paper 2 presents an adaptive PolSAR superpixel generation method based on the simple linear iterative clustering (SLIC) framework. In the k-means clustering procedure, multiple cues including polarimetric, spatial, and texture information are considered to measure the distance. Since the constant weighting factor which balances the spectral similarity and spatial proximity may cause over- or under-superpixel segmentation in different areas, the proposed method sets the factor adaptively based on the homogeneity analysis. Then, in heterogeneous areas, the spectral similarity is more significant than the spatial constraint, generating superpixels which better preserved local details and refined structures. Paper 3 investigates another PolSAR superpixel generation method, which is achieved from the global optimization aspect, using the entropy rate method. The distance between neighbouring pixels is calculated based on their corresponding DSDA regions. In addition, the SIRV distance and the Wishart distance are combined together. Therefore, the proposed method makes good use of the entropy rate framework, and also incorporates the merits of the SIRV distance and the Wishart distance. The superpixels are generated in a homogeneity-adaptive manner, resulting in smooth representation of the land covers in homogeneous areas, and well preserved details in heterogeneous areas. / <p>QC 20171123</p>