Adaptive multiscale feature extraction from range data

We present a method to extract meaningful features from range images which introduces some novel ideas: First, our primitives are a richer set than the usual one, as we not only extract high frequences events, which correspond to jump boundaries and sharp creases, but also low frequency events such as smooth ridges and ravines. These last features are generally discarded as they tend to hide in noise; however, they provide a coarse description of the shape, and we believe they may help in inferring volumetric descriptions. The method works at multiple scales to improve reliability, and we have designed a control strategy to automaticaNy select the most appropriate mask size. We present results on images of different levels of complexity from three different sensors.