A Parallel Version for the Propagation Algorithm ⋆ M´ arcio Bastos Castro Lucas Baldo Luiz Gustavo Fernandes Mateus Raeder Pedro Velho ProgramadeP´os-Gradua¸c˜aoemCiˆ enciadaComputa¸c˜ao,PUCRS Avenida Ipiranga, 6681 - CEP 90619-900, Porto Alegre, Brazil {mcastro, lbaldo, gustavo, mraeder, pedro}@inf.pucrs.br Abstract. This paper presents a parallel version for the Propagation Algorithm which belongs to the region growing family of algorithms. The main goal of our implementation is to decrease de Propagation Al- gorithm execution time in order to allow its use on image interpolation applications. Our solution is oriented to low cost high performance plat- forms such as clusters of workstations. Four different input data sets represented by pairs of images were chosen in order to carry out experi- mental tests. The results obtained show that our parallel version of the Propagation Algorithm presents significant speedups. 1 Introduction Creating virtual in-between views from two scenes of the same subject taken from different points of view can be a very interesting tool to economize re- sources in some practical applications [1]. One main example is typically found in teleconferencing with limited network bandwidth. Image-based interpolation is a method to create smooth and realistic virtual views between two original view points. Interpolation applications are usually based on a three-phase algo- rithm [2]: construction of a dense matching map between the original images, separations of matched areas from unmatched ones and finally the generation of all in-between images. The matching phase is by far the most time consuming one of this procedure. The general technique for matching areas from different images is called region growing. Its basic principle is the use of images charac- teristics to group neighbor pixels and thus creating regions. In [3], a new region growing algorithm was proposed. It is based on the construction of a quasi-dense matching map between the two original views and it is able to perform more ac- curate matches. Its originality consists on the adoption of a “best first” strategy to select the next match from a set of seed matches which is updated through the addition of each new found match from the precedent algorithm iteration. This new algorithm was called the Propagation Algorithm, and the improvements on the matching procedure brought together an additional computational cost. This paper proposes a parallel version for the Propagation Algorithm. The target ar- chitecture is a cluster of workstations and the implementation was carried out using the standard message passing library MPI [4]. ⋆ This work was developed in collaboration with HP Brazil R&D.