Robust Multi-Structure Vision Data Segmentation: Local Optimisation vs Random Sampling
Mohammad Reza Alizadeh, Alireza Bab‐Hadiashar, Reza Hoseinnezhad, Zhenwei Cao · Swinburne figshare (Swinburne University of Technology) · 2011
Many applications in computer vision and robotics involve analysing acquired vision data to recognize and extract known patterns embedded in data. There have been various robust model-fitting and segmentation techniques developed for applications such as motion estimation and range segmentation. This paper presents a robust model fitting and clustering method, which is an improvement over the recently published Jlinkage method. The main contribution of this paper is the elimination of a pair of manually tuned parameters in the Jlinkage method, as the presence of those parameters results in a cumbersome trial-and-error process to find the best initial values suitable for specific data. To achieve this, the first phase of the Jlinkage method was modified using a local optimisation scheme to generate hypotheses, instead of random sampling. Moreover, a data-driven scale estimator and segmentation method was used to find a consensus set of hypotheses instead of using the threshold parameter. The resulting method was evaluated through several experiments on the synthetic multi-structure 2D and 3D data and real-world 3D range data segmentation application. Results show the capability and flexibility of the presented method and its improved performance compared to the J-linkage technique.