Publicação

Estruturas circulantes em visão por computador

Ver documento

Detalhes bibliográficos
Resumo:Visual recognition systems are extremely data-hungry. To accurately recognize a new kind of object, a learning algorithm requires a massive dataset of example images, often augmented artificially by cropping different image regions. More examples seem to invariably raise the computational burden of learning. Is this an inescapable fact? In this thesis, we show that it is not true -- that the structure of these datasets hides important shortcuts. The key observation is that samples are not independent, since samples cropped from the same image share most pixels. Using an analytical model of image translation, the cyclic shift, we show that the resulting dataset contains circulant matrices. As a result, we can diagonalize it with the Discrete Fourier Transform (DFT), which reduces both storage and computation by orders of magnitude. The use of the DFT further reveals an interesting link to correlation filters from classical signal processing. We accelerate learning algorithms such as Ridge Regression and Support Vector Regression, addressing linear and non-linear kernel methods. We propose two trackers, the Dual and Kernelized Correlation Filters, which run at hundreds of frames-per-second, and yet perform better than more complex trackers on a 50 videos benchmark. For detection, we propose a decomposition that is several times faster than hard-negative mining, a staple of detector learning. We also generalize these results for other kinds of datasets, such as rotated images or non-rigidly deformed images, which accelerates the learning of pose estimators. The proposed solutions require only a few lines of code to implement, relying on the Fast Fourier Transform and optional off-the-shelf solvers for the bulk of the computations, which easily run in parallel. The software produced during this thesis is open-source.
Autores principais:Henriques, João Filipe dos Santos
Outros Autores:Henriques, João Filipe dos Santos
Assunto:Circulant Matrices Discrete Fourier Transform Correlation Filters Image Transformations Visual Tracking Object Detection Pose Estimation
Ano:2016
País:Portugal
Tipo de documento:tese de doutoramento
Tipo de acesso:acesso aberto
Instituição associada:Universidade de Coimbra
Idioma:inglês
Origem:Estudo Geral - Universidade de Coimbra
Descrição
Resumo:Visual recognition systems are extremely data-hungry. To accurately recognize a new kind of object, a learning algorithm requires a massive dataset of example images, often augmented artificially by cropping different image regions. More examples seem to invariably raise the computational burden of learning. Is this an inescapable fact? In this thesis, we show that it is not true -- that the structure of these datasets hides important shortcuts. The key observation is that samples are not independent, since samples cropped from the same image share most pixels. Using an analytical model of image translation, the cyclic shift, we show that the resulting dataset contains circulant matrices. As a result, we can diagonalize it with the Discrete Fourier Transform (DFT), which reduces both storage and computation by orders of magnitude. The use of the DFT further reveals an interesting link to correlation filters from classical signal processing. We accelerate learning algorithms such as Ridge Regression and Support Vector Regression, addressing linear and non-linear kernel methods. We propose two trackers, the Dual and Kernelized Correlation Filters, which run at hundreds of frames-per-second, and yet perform better than more complex trackers on a 50 videos benchmark. For detection, we propose a decomposition that is several times faster than hard-negative mining, a staple of detector learning. We also generalize these results for other kinds of datasets, such as rotated images or non-rigidly deformed images, which accelerates the learning of pose estimators. The proposed solutions require only a few lines of code to implement, relying on the Fast Fourier Transform and optional off-the-shelf solvers for the bulk of the computations, which easily run in parallel. The software produced during this thesis is open-source.