Multi-resolution binary shape tree for efficient 2D clustering

The analysis of discrete two-dimensional distributions is a relevant task in computer vision, since many intermediate representations are generated inform of a two-dimensional map. Probabilistic inference or the response of discriminative classification often yield multi-modal distributions in form of 2D digital images, where the accurate and computationally efficient delineation of structures with varying attributes such as scale, orientation and shape represents a challenge. The simplest example is non-maximum suppression, where typically the response of a center-surround structural element applied as a filter is used to suppress spurious detection responses. In this paper we propose a simple scheme which is capable to delineate the shape of arbitrary distributions around a local density maximum driven by a local binary shape model, resulting in consistent object hypotheses. We employ a coarse-to-fine analysis scheme where learned binary shapes of increasing resolution guide a shape matching process. We demonstrate applicability for delineating compact clusters in a noisy probabilistic occupancy map, and the capability for detecting structurally consistent line structures in a text detector response map. Results are compared to other spatial grouping schemes and obtained results demonstrate a fast and accurate delineation performance.