Role of scale in partitioning shape

We present three distinct improvements to the neck and limb-based approach to partitioning of visual form, where pairs of negative curvature minima with good continuation of their respective tangents constitute limbs, and locally shortest lines through the shape constitute necks. The numerous conflicting limb and neck hypotheses are resolved through a notion of salience, leading to intuitive parts for smooth shapes. We improve on this approach in three ways. First a significant difficulty is dealing with scale, e.g., in partitioning noisy shapes where some perceptually valid limbs are not marked by salient negative curvature minima, and where numerous hypotheses arise due to noise. We present an approach which can reliably construct and detect coarse-scale negative curvature extremas. Second, we devise a multiscale partitioning scheme, thus minimizing the coarse-scale interaction of part-lines arising from features at different scales. Third, we employ an Euler spiral as the part-curve instead of a straight part-line for partitioning shape thus improving the partitioning results.