Abstract
This paper presents a novel method to lo-calize and segment objects on close-range table-topscenarios sensed with a depth sensor. The method isbased on a novelobjectnessmeasure that evaluateshow likely a 3D region in space (defined by an ori-ented bounding box) could contain an object. Withina parametrized volume of interest placed above thetable plane, a set of 3D bounding boxes is generatedthat exhaustively covers the parameter space. Effi-ciently evaluating - thanks to integral volumes andparallel computing - the 3D objectness at each sam-pled bounding box allows efficiently defining a setof regions in space with high probability of contain-ing an object. Bounding boxes characterized by highobjectness are then processed by means of a globaloptimization stage aimed at discarding inconsistentobject hypotheses with respect to the scene. We eval-uate the effectiveness of the method for the task ofscene segmentation
Reference
Aldoma, A., Tombari, F., Kropatsch, W., & Vincze, M. (2013). Localizing and Segmenting Objects with 3D Objectness. In W. Kropatsch, F. Torres Garcia, & G. Ramachandran (Eds.), Proceedings of the 18th Computer Vision Winter Workshop 2013 (pp. 86–93). Prip 186/3. http://hdl.handle.net/20.500.12708/54742