Interest point detectors are important components in a variety of computer vision systems. This paper demonstrates an automated virtual 3D environment for controlling and measuring detected interest points on 2D images in an accurate and rapid manner. Real-time affine transform tools enable easy implementation and full automation of complex scene evaluations without the time-cost of a manual setup. Nine detectors are tested and compared using evaluation and testing methods based on Schmid, Mohr, and Bauckhage (2000). Each detector is tested on the BSDS500 image set and 34 3D scanned, and manmade models using rotation in the X, Y, and Z axis as well as scale in the X,Y axis. Varying degrees of noise on the models are also tested. Results demonstrate the differing performance and behaviour of each detector across the evaluated transformations, which may assist computer vision practitioners in choosing the right detector for their application.
Interest Point Detectors Stability Evaluation on ApolloScape Dataset
