Citation

Eng, Wei Yong (2020) Improved Precision Dense Descriptor-Based Matching For Large Displacement Flow. PhD thesis, Multimedia University.

Full text not available from this repository.

Abstract

Optical flow is a traditional problem in computer vision. It is often formulated elegantly in a Markov Random Formulation and divided into two algorithms: i) local method with data term and ii) global method with both data and smoothness terms. In a local method, it is fuelled by the patch match to find the most similar patch between input image pairs. The corresponding patch pairs are sought for all the patches in a reference image within a search range. The exhaustive search for the corresponding pair is computationally expensive. Previous literature focused on reducing the computational load and removing its dependency on the window size and search range, during the search process. Inspired by the PatchMatch and guided filter, PatchMatch Filter explored the possibility of a simultaneous computational reduction in both search range and window size dimensions. Most research works focus on solving small displacement optical flow problem as it involves mostly translational movement and little or no scale and orientation changes. Recently more and more researchers have explored large displacement optical flow as it is closer to a realistic environment and of more general usage. Large displacement optical flow is more challenging in that it involves significant geometric distortion such as scale and orientation changes. Pixel-based matching method is typically used to solve small displacement optical flow and it has accurate spatial precision as well as it is computationally efficient. Some research works adopted descriptor-based matching sparsely to deal with the significant scale and orientation changes, in which pure pixel-based matching might fails. Descriptor-based matching is known to deal with geometric variation at the cost of spatial resolution, which is an inherent limitation of the histogram computation nature. This work proposes to adopt non-histogram based features like pixel colour and gradient into typical histogram-based feature matching to obtain high precision flow. In the proposed method, the label space is parameterised in an affine motion model which includes scaling, orientation, horizontal and vertical translations. Also, Daisy descriptor is employed to deal with geometric changes.

Item Type:	Thesis (PhD)
Additional Information:	Call No: TA1630 .E54 2020
Uncontrolled Keywords:	Optical data processing
Subjects:	T Technology > TA Engineering (General). Civil engineering (General) > TA1501-1820 Applied optics. Photonics
Divisions:	Faculty of Engineering and Technology (FET)
Depositing User:	Ms Nurul Iqtiani Ahmad
Date Deposited:	19 May 2023 08:30
Last Modified:	19 May 2023 08:30
URII:	http://shdl.mmu.edu.my/id/eprint/11411

Downloads

Edit (login required)