A comparison of interest point and region detectors on structured, range and texture images

Abstract

This article presents an evaluation of the image retrieval and classification potential of local features. Several affine invariant region and scale invariant interest point detectors in combination with well known descriptors were evaluated. Tests on building, range and texture databases were carried out in order to understand the effects of the nature and the variability of the data on the performance of the detectors in terms of their invariance to affine deformations and scale changes. Furthermore, a novel multi-scale edge shape detector, Twin Leaf Regions (TLR) is also proposed using a graph based image decomposition. In TLR, Affine adaptation is avoided in order to reduce the offset from the edges so that pure edges shapes are captured in multiple scales. In the evaluation of building recognition, both homogeneous affine regions (such as Maximally Stable Extremal Regions (MSER)) and corner based detectors (such as Hessian and Harris with both Affine/Laplace variants, SURF with determinant of Hessian based corners and SIFT with difference of Gaussians) acquired more than 90% mean average precision, whereas on range images, homogeneous region detector did not work well. TLR offered good performance than MSER and comparable performance to Harris Affine and Harris Laplace in range image classification and texture retrieval. But its performance was low in building recognition. In general, it was observed that the affine and scale invariance becomes less effective in range and textured images. It is also shown that in a bi-channel approach, combining surface and edge regions (MSER and TLR) boosts the overall performance. Among the descriptors, SIFT and SURF generally offer higher performance but low dimensional descriptors such as Steerable Filters follow closely.