Evaluation of Low-Level Image Representations for Illumination-Insensitive Recognition of Textureless Objects

In this paper the problem of recognizing textureless objects in unconstrained illumination and material conditions is investigated. We evaluate the discriminative power of various low-level image features for a pixelwise representation of the underlying surface characteristics of the object. For this purpose, a new dataset with rendered images of 3D models is used which allows to directly compare the influences of texture and material properties in an object recognition scenario. The results are further validated on a dataset of real object images and finally reveal that jets of singleand multi-scale even Gabor filter responses outperform other proposed features in scenarios with textureless objects and strong variations of illumination.