Evaluating Colour-Based Object Recognition Algorithms Using the SOIL-47 Database

In this paper, a new image set, called the Surrey Object Image Library (SOIL-47) is introduced, on which the performance of two colour-based object recognition methods is evaluated. The data was collected specifically for testing colour-based recognition algorithm and is publicly available. In the conducted experiments on SOIL-47, we evaluate two recognition algoritms; the Multimodal Neighbourhood Signature (MNS) approach and a method based on a Attributed Relational Graph (ARG). The MNS approach represents object appearance by measurements computed from image neighbourhoods with a multimodal colour density function. The ARG approach computes a graph of affine invariant measurements of the colour and shape of segmented image regions. Using only a single model image of each of the 47 objects, MNS performed well even for extreme test views close to 90 degrees. The ARG method assumes a locally planar surface, therefore a second experiment was conducted on a subset of box-like objects of SOIL-47. MNS performance was fairly stable, outperforming ARG for most viewing angles.. Note, that this is the first systematic test of MNS with controlled 3D viewpoint change.