In this paper, a new technique for 3D acoustic image segmentation and modelling is proposed. Especially, in the underwater environment, in which optical sensors suffer from visibility problems, the acoustical devices may provide efficient solutions, but, on the other hand, acoustic image interpretation is surely more difficult for a human operator. The proposed application involves the use of an acoustic camera which directly acquires images structured as a set of 3D points. Due to the noisy nature of this type of data, the segmentation problem becomes more challenging and the standard algorithms for range image segmentation are likely to fail. The proposed method is based on a simplified version of the so called recover and select paradigm in which the seed areas, from which the segmentation starts, are generated by adopting a robust approach based on the RANSAC (RANdom Sample And Consensus) algorithm. Superquadric primitives are directly recovered from raw data without any pre-segmentation processing. Experimental trials using both synthetic and real acoustical images confirm the goodness of the method, and a large robustness of the resulting segmented images, associated to a relatively low computational load.