Distinctive texture features from perspective-invariant keypoints

In this paper, we present an algorithm to detect and describe features of surface textures, similar to SIFT and SURF. In contrast to approaches solely based on the intensity image, it uses depth information to achieve invariance with respect to arbitrary changes of the camera pose. The algorithm works by constructing a scale space representation of the image which conserves the real-world size and shape of texture features. In this representation, keypoints are detected using a Difference-of-Gaussian response. Normal-aligned texture descriptors are then computed from the intensity gradient, normalizing the rotation around the normal using a gradient histogram. We evaluate our approach on a dataset of planar textured scenes and show that it outperforms SIFT and SURF under large viewpoint changes.