Using range information to reduce local descriptor computations: A case study

Robot vision systems commonly use advanced image processing techniques, such as the well-known SIFT local descriptors, for object recognition. While SIFT descriptors provide excellent recognition power, the processing time required for generating the descriptors is significant. We therefore explore two methods for utilizing range images to improve the performance of local descriptor computations. The first method involves segmenting the input image based on range information, to focus local descriptor computation to relevant parts of the image. The second method uses range information to reduce the effort needed to maintain scale-invariance. Our results indicate that both techniques can significantly reduce computation cost while preserving recognition quality.