Streamed watershed transform on GPU for processing of large volume data

Since its introduction the watershed transform became a popular method for volume data segmentation. A range of various algorithms for its computation were developed, including parallel algorithms for computation on different architectures. Recently also algorithms for consumer graphical accelerators were developed. Neither of these, however, are able to process data larger than the available memory as the whole data has to be present in the memory of the device. In this paper we present two versions of a streamed multi-pass algorithm for watershed computation on a GPU. As the slice-based streaming approach is used both variants are capable of processing data exceeding the size of the available graphics accelerator memory.