Repeatability Measurements for 2D Interest Point Detectors on 3D Models

Abstract

Interest point detectors typically operate on 2D images, yet these frequently constitute projections of real 3D scenes [8]. Analysing and comparing the performance of these detectors as to their utility at tracking points in a 3D space is challenging. This paper demonstrates a virtual 3D environment which can measure the repeatability of detected interest points accurately and rapidly. Real-time 3D transform tools enable easy implementation of complex scene evaluations without the time-cost of a manual setup or mark-up. Nine detectors are tested and compared using evaluation and testing methods based on Schmid [16]. Each detector is tested on 34 textured and untextured models that are either scanned from physical objects or modelled by an artist. Rotation in the X, Y, and Z axis as well as scale transformations are tested on each model, with varying degrees of artificial noise applied. Results demonstrate the performance variability of different interest point detectors under different transformations and may assist researchers in deciding on the correct detector for their computer vision application.