Stochastic Motion Clustering

摘要: This paper presents a new method for motion segmentation, the clustering together of features that belong to independently moving objects. The exploits fact two views rigidly connected 3D point set are linked by 3×3 Fundamental Matrix which contains all information on given correspondences. segmentation problem is transformed into one finding Matrices optimally describe observed temporal correspondences, where optimization couched as maximization posteriori probability an interpretation data. To reduce search space, feasible clusters hypothesized using robust statistical techniques, and multiple hypothesis test performed determine particular combination many most likely represent actual feature motions observed. shown be computable in terms 0–1 integer programming method, alleviating combinatorial computing difficulties inherent such problems.