|
|
Accelerated Sparse Matrix Operations in Nonlinear Least Squares Solvers
Polok, Lukáš ; Hartley, Richard (oponent) ; Sojka, Eduard (oponent) ; Smrž, Pavel (vedoucí práce)
This thesis focuses on data structures for sparse block matrices and the associated algorithms for performing linear algebra operations that I have developed. Sparse block matrices occur naturally in many key problems, such as Nonlinear Least Squares (NLS) on graphical models. NLS are used by e.g. Simultaneous Localization and Mapping (SLAM) in robotics, Bundle Adjustment (BA) or Structure from Motion (SfM) in computer vision. Sparse block matrices also occur when solving Finite Element Methods (FEMs) or Partial Differential Equations (PDEs) in physics simulations. The majority of the existing state of the art sparse linear algebra implementations use elementwise sparse matrices and only a small fraction of them support sparse block matrices. This is perhaps due to the complexity of sparse block formats which reduces computational efficiency, unless the blocks are very large. Some of the more specialized solvers in robotics and computer vision use sparse block matrices internally to reduce sparse matrix assembly costs, but finally end up converting such representation to an elementwise sparse matrix for the linear solver. Most of the existing sparse block matrix implementations focus only on a single operation, such as the matrix-vector product. The solution proposed in this thesis covers a broad range of functions: it includes efficient sparse block matrix assembly, matrix-vector and matrix-matrix products as well as triangular solving and Cholesky factorization. These operations can be used to construct both direct and iterative solvers as well as to compute eigenvalues. Highly efficient algorithms for both Central Processing Units (CPUs) and Graphics Processing Units (GPUs) are provided. The proposed solution is integrated in SLAM++ , a nonlinear least squares solver focused on robotics and computer vision. It is evaluated on standard datasets where it proves to significantly outperform other similar state of the art implementations, without sacrificing generality or accuracy in any way.
|
|
|
Acceleration of Object Detection Using Classifiers
Juránek, Roman ; Kälviäinen, Heikki (oponent) ; Sojka, Eduard (oponent) ; Zemčík, Pavel (vedoucí práce)
Detection of objects in computer vision is a complex task. One of most popular and well explored approaches is the use of statistical classifiers and scanning windows. In this approach, classifiers learned by AdaBoost algorithm (or some modification) are often used as they achieve low error rates, high detection rates and they are suitable for detection in real-time applications. Object detection run-time which uses such classifiers can be implemented by various methods and properties of underlying architecture can be used for speed-up of the detection. For the purpose of acceleration, graphics hardware, multi-core architectures, SIMD or other means can be used. The detection is often implemented on programmable hardware. The contribution of this thesis is to introduce an optimization technique which enhances object detection performance with respect to an user defined cost function. The optimization balances computations of previously learned classifiers between two or more run-time implementations in order to minimize the cost function. The optimization method is verified on a basic example -- division of a classifier to a pre-processing unit implemented in FPGA, and a post-processing unit in standard PC.
|
|
|
Image Restoration Based on Convolutional Neural Networks
Svoboda, Pavel ; Baláž, Teodor (oponent) ; Sojka, Eduard (oponent) ; Zemčík, Pavel (vedoucí práce)
A merit of this thesis is to introduce a unified image restoration approach based on a convolutional neural network which is to some degree degradation type independent. Convolutional neural network models were trained for two different tasks, a motion deblurring of license plate images and a removal of artifacts related to lossy image compression. The capabilities of such models are studied from two main perspectives. Firstly, how well the model can restore an image compared to the state-of-the-art methods. Secondly, what is the model's ability to handle several ranges of the same degradation type. An idea of the unified end-to-end approach is based on a recent development of neural networks and related deep learning in a field of computer vision. The existing hand-engineered methods of image restoration are often highly specialized for a given degradation type and in fact, define state of the art in several image restoration tasks. The end-to-end approach allows to directly train the required model on specifically corrupted images, and, further, to restore various levels of corruption with a single model. For motion deblurring, the end-to-end mapping model derived from models used in computer vision is deployed. Compression artifacts are restored with similar end-to-end based model further enhanced using specialized objective functions together with a network skip architecture. A direct comparison of the convolutional network based models and engineered methods shows that the data-driven approach provides beyond state-of-the-art results with a high ability to generalize over different levels of degradations. Based on the achieved results, this work presents the convolutional neural network based methods suggesting a possibility having the unified approach used for wide range of image restoration tasks.
|
|
|
HUMAN ACTION RECOGNITION IN VIDEO
Řezníček, Ivo ; Baláž, Teodor (oponent) ; Sojka, Eduard (oponent) ; Zemčík, Pavel (vedoucí práce)
This thesis focuses on the improvement of human action recognition systems. It reviews the state-of-the-art in the field of action recognition from video. It describes techniques of digital image and video capture, and explains computer representations of image and video. This thesis further describes how local feature vectors and local space-time feature vectors are used, and how captured data is prepared for further analysis, such as classification methods. This is typically done with video segments of arbitrarily varying length. The key contribution of this work explores the hypothesis that the analysis of different types of actions requires different segment lenghts to achieve optimal quality of recognition. An algorithm to find these optimal lengths is proposed, implemented, and tested. Using this algorithm, the hypothesis was experimentally proven. It was also shown that by finding the optimal length, the prediction and classification power of current algorithms is improved upon. Supporting experiments, results, and proposed exploitations of these findings are presented.
|
|
|
Identifikace vozidel na snímcích dopravních situací
Petyovský, Petr ; Sojka, Eduard (oponent) ; Železný, Miloš (oponent) ; Horák, Karel (vedoucí práce)
Cílem této disertační práce je návrh metod pro získání dalších parametrů o vozidle ze snímků reálné dopravní situace ke stávající informaci o RZ vozidla a jeho poloze v měřeném úseku. Úkolem je využít stávající instalace kamerových systémů a na základě dat získaných z těchto zařízení navrhnout nové metody extrakce dalších parametrů o vozidle. Řešení lze rozdělit na dvě skupiny: 1. Metody pro získání příznaků a metody vyhodnocení dat, které povedou k rozpoznání typu vozidla na základě jediného snímku vozidla. 2. Metody pro získání údajů o tvaru vozidla na základě sekvence snímků projíždějícího vozidla
|
|
|
Lifting Scheme Cores for Wavelet Transform
Bařina, David ; Kälviäinen, Heikki (oponent) ; Sojka, Eduard (oponent) ; Zemčík, Pavel (vedoucí práce)
The thesis focuses on efficient computation of the two-dimensional discrete wavelet transform. The state-of-the-art methods are extended in several ways to perform the transform in a single loop, possibly in multi-scale fashion, using a compact streaming core. This core can further be appropriately reorganized to target the minimization of certain platform resources. The approach presented here nicely fits into common SIMD extensions, exploits the cache hierarchy of modern general-purpose processors, and is suitable for parallel evaluation. Finally, the approach presented is incorporated into the JPEG 2000 compression chain, in which it has proved to be fundamentally faster than widely used implementations.
|
|
|
Acceleration of 2D Wavelet transform on parallel architectures
Kula, Michal ; Schier, Jan (oponent) ; Sojka, Eduard (oponent) ; Zemčík, Pavel (vedoucí práce)
Although a 2D discrete wavelet transform has been widely studied during the last two decades, some directions were not examined from all points of view. One of these directions is a technique for calculating such transform that has various balanced barriers, arithmetic operations, and memory usage focused on various architectures. This thesis shows several new methods of calculation of such transform with variously balanced operations. These methods are widely described and their behaviour is evaluated on several graphics adapters using GPGPU, graphics pipeline, and multicore CPU architectures using OpenMP.
|
|
|
Recognition of Partially Occluded Objects
Krolupper, Filip ; Flusser, Jan (vedoucí práce) ; Sojka, Eduard (oponent) ; Peters, Gabriele (oponent)
In this thesis we focus on partially occluded object recognition under geometric transformations. Objects are represented by their contours. Depending on the kind of geometric transformation and robustness to occlusion we introduce different solutions. Our results are applicable in industry, robotics, 3D vision, forensics, etc. We propose three novel methods for partially occluded object recognition. The major contribution of all our methods is a creation of features. Features are designed to be local and invariant to appropriate geometric transformations. We use mostly standard feature matching to prove properties of designed features. The first method deals only with translation, rotation and scaling (Euclidian transformation) and is based on contour approximation by circle arcs. The parameters of the circle arcs seem to be suitable features. The second method deals with affine transformation and is based on polygonal approximation of contours and, moreover, is robust to additive noise. The second method splits the contour into parts using inflexion points and transforms every part into both normalized shape and position. The parameters of standard shapes of every part are the desired features. The third method deals also with affine transformation. It splits the object into parts using a novel, cutting...
|
|
|
Vehicle Speed Measurement Using Stereo Camera Pair
Najman, Pavel ; Sojka, Eduard (oponent) ; Guillemaut, Jean-Yves (oponent) ; Zemčík, Pavel (vedoucí práce)
This thesis aims to answer the question whether it is currently possible to autonomously measure the speed of vehicles using a stereoscopic method with the average error within 1 km/h, the maximum error within 3 km/h, and the standard deviation within 1 km/h. The error ranges are based on the requirements of the OIML whose Recommendations serve as templates for metrological legislations of many countries. To answer this question, a~hypothesis is formulated and tested. A method that utilizes a stereo camera pair for vehicle speed measurement is proposed and experimentally evaluated. The experiments show that the technique overcomes state-of-the-art results with the mean error of approximately 0.05 km/h, the standard deviation of less than 0.20 km/h, and the maximum absolute error of less than 0.75 km/h. The results are within the required ranges, and therefore the formulated hypothesis holds.
|
|
|
Accelerated Sparse Matrix Operations in Nonlinear Least Squares Solvers
Polok, Lukáš ; Hartley, Richard (oponent) ; Sojka, Eduard (oponent) ; Smrž, Pavel (vedoucí práce)
This thesis focuses on data structures for sparse block matrices and the associated algorithms for performing linear algebra operations that I have developed. Sparse block matrices occur naturally in many key problems, such as Nonlinear Least Squares (NLS) on graphical models. NLS are used by e.g. Simultaneous Localization and Mapping (SLAM) in robotics, Bundle Adjustment (BA) or Structure from Motion (SfM) in computer vision. Sparse block matrices also occur when solving Finite Element Methods (FEMs) or Partial Differential Equations (PDEs) in physics simulations. The majority of the existing state of the art sparse linear algebra implementations use elementwise sparse matrices and only a small fraction of them support sparse block matrices. This is perhaps due to the complexity of sparse block formats which reduces computational efficiency, unless the blocks are very large. Some of the more specialized solvers in robotics and computer vision use sparse block matrices internally to reduce sparse matrix assembly costs, but finally end up converting such representation to an elementwise sparse matrix for the linear solver. Most of the existing sparse block matrix implementations focus only on a single operation, such as the matrix-vector product. The solution proposed in this thesis covers a broad range of functions: it includes efficient sparse block matrix assembly, matrix-vector and matrix-matrix products as well as triangular solving and Cholesky factorization. These operations can be used to construct both direct and iterative solvers as well as to compute eigenvalues. Highly efficient algorithms for both Central Processing Units (CPUs) and Graphics Processing Units (GPUs) are provided. The proposed solution is integrated in SLAM++ , a nonlinear least squares solver focused on robotics and computer vision. It is evaluated on standard datasets where it proves to significantly outperform other similar state of the art implementations, without sacrificing generality or accuracy in any way.
|
host ::
RSS.