Faster and better: a machine learning approach to corner detection

arXiv:0810.2434v1 [cs.CV] 14 Oct 2008 1 Faster and better: a machine learning approach to corner detection Edward Rosten, Reid Porter, and Tom Drummond Edward Rosten and Reid Porter are with Los Alamos National Laboratory, Los Alamos, New Mexico, USA, 87544. Email: edrosten@lanl.gov, rporter@lanl.gov Tom Drummond is with Cambridge University, Cambridge University Engineering Department, Trumpington Street, Cam- bridge, UK, CB2 1PZ Email: twd20@cam.ac.uk November 26, 2024 DRAFT 2 Abstract The repeatability and efficiency of a corner detector determines how likely it is to be useful in a real-world application. The repeatability is importand because the same scene viewed from different positions should yield features which correspond to the same real-world 3D locations [1]. The efficiency is important because this determines whether the detector combined with further processing can operate at frame rate. Three advances are described in this paper. First, we present a new heuristic for feature detection, and using machine learning we derive a feature detector from this which can fully process live PAL video using less than 5% of the available processing time. By comparison, most other detectors cannot even operate at frame rate (Harris detector 115%, SIFT 195%). Second, we generalize the detector, allowing it to be optimized for repeatability, with little loss of efficiency. Third, we carry out a rigorous comparison of corner detectors based on the above repeatability criterion applied to 3D scenes. We show that despite being principally constructed for speed, on these stringent tests, our heuristic detector significantly outperforms existing feature detectors. Finally, the comparison demonstrates that using machine learning produces significant improvements in repeatability, yielding a detector that is both very fast and very high quality. Index Terms Corner detection, feature detection. I. INTRODUCTION Corner detection is used as the first step of many vision tasks such as tracking, localisation, SLAM (simultaneous localisation and mapping), image matching and recognition. This need has driven the development of a large number of corner detectors. However, despite the massive increase in computing power since the inception of corner detectors, it is still true that when processing live video streams at full frame rate, existing feature detectors leave little if any time for further processing. In the applications described above, corners are typically detected and matched into a database, thus it is important that the same real-world points are detected repeatably from multiple views [1]. The amount of variation in viewpoint under which this condition should hold depends on the application. November 26, 2024 DRAFT 3 II. PREVIOUS WORK A. Corner detectors Here we review the literature to place our advances in context. In the literature, the terms “point feature”, “feature”, “interest point” and “corner” refer to a small point of interest with variation in two dimensions. Such points often arise as the result of geometric discontinuities, such as the corners of real world objects, but they may also arise from small patches of texture. Most algorithms are capable of detecting both kinds of points of interest, though the algorithms are often designed to detect one type or the other. A number of the detectors described below compute a corner response, C, and define corners to be large local maxima of C. 1) Edge based corner detectors: An edge (usually a step change in intensity) in an image corresponds to the boundary between two regions. At corners, this boundary changes direction rapidly. a) Chained edge based corner detectors: Many techniques have been developed which involved detecting and chaining edges with a view to analysing the properties of the edge, often taking points of high curvature to be corners. Many early methods used chained curves, and since the curves are highly quantized, the techniques concentrate on methods for effectively and efficiently estimating the curvature. A common approach has been to use a chord for estimating the slope of a curve or a pair of chords to find the angle of the curve at a point. Early methods computed the smallest angle of the curve over chords spanning different numbers of links. Corners are defined as local minima of angle [2] after local averaging [3]. Alternatively, corners can be defined as isolated discontinuities in the mean slope, which can be computed using a chord spanning a fixed set of links in the chain [4]. Averaging can be used to compute the slope and the length of the curve used to determine if a point is isolated [5]. The angle can be computed using a pair of chords with a central gap, and peaks with certain widths (found by looking for zero crossings of the angle) are defined as corners [6]. Instead of using a fixed set of chord spans, some methods compute a ‘region of support’ which depends on local curve properties. For instance local maxima of chord lengths

…(Full text truncated)…

This content is AI-processed based on ArXiv data.