Robust multi-model fitting via neighborhood graph structure consistency

Recently, some graph-based methods have been proposed for multi-model fitting. These methods usually construct full-connected graphs between any pairs of data for model fitting. However, the constructed graphs tend to contain many invalid edges, especially for input data corrupted by a large number of noise and outliers, which may lead to suboptimal fitting results. In this paper, we propose a novel multi-model fitting method based on neighborhood graph structure consistency (NGSC), which preserves the local neighborhood structures of potential inliers to construct an effective graph for robust model fitting. Specifically, by exploiting the motion and preference information of input data, we adaptively generate the local-neighbors for each vertex of the graph. To enhance the effectiveness of graph construction, we propose an effective guided sampling strategy to improve the accuracy of preference correlation measurement. In addition, we propose a novel graph clustering algorithm to effectively detect the connected components of the graph (i.e., connected subgraphs) corresponding to model instances, for model selection. Overall, the key elements (i.e., graph construction, graph partition, and subgraph detection) are original. Extensive experimental results on several challenging datasets show that the proposed NGSC achieves more accurate fitting results than other several state-of-the-art fitting methods.