Localizing and segmenting objects with 3D objectness

This paper presents a novel method to localize and segment objects on close-range table-top scenarios acquired with a depth sensor. The method is based on a novel objectness measure that evaluates how likely a 3D region in space (defined by an oriented 3D bounding box) could contain an object. Within a parametrized volume of interest placed above the table plane, a set of 3D bounding boxes is generated that exhaustively covers the parameter space. Efficiently evaluating — thanks to integral volumes and parallel computing— the 3D objectness at each sampled bounding box allows defining a set of regions in space with high probability of containing an object. Bounding boxes characterized by high objectness are then processed by means of a global optimization stage aimed at discarding inconsistent object hypotheses with respect to the scene. We evaluate the effectiveness of the method for the task of scene segmentation.