System-theoretic approach to image interest point detection

TL;DR

This paper introduces a system-theoretic framework for interest point detection in images, focusing on reducing computational complexity by analyzing detector-descriptor interdependencies and creating optimized, irredundant detectors.

Contribution

It presents a novel system-theoretic approach that identifies detector redundancy and constructs optimized detectors with lower computational costs.

Findings

Irredundant detectors have lower computational complexity.

The approach generalizes existing methods for reducing image registration complexity.

A systematic analysis of detector-descriptor interdependency is provided.

Abstract

Interest point detection is a common task in various computer vision applications. Although a big variety of detector are developed so far computational efficiency of interest point based image analysis remains to be the problem. Current paper proposes a system-theoretic approach to interest point detection. Starting from the analysis of interdependency between detector and descriptor it is shown that given a descriptor it is possible to introduce to notion of detector redundancy. Furthermore for each detector it is possible to construct its irredundant and equivalent modification. Modified detector possesses lower computational complexity and is preferable. It is also shown that several known approaches to reduce computational complexity of image registration can be generalized in terms of proposed theory.