Dynamic spatio-temporal modeling for example-based human silhouette recovery

In this paper, we pose human silhouette recovery as a problem of robust spatio-temporal signal restoration, which aims to effectively recover the original human silhouette signals from noisy corruption or partial occlusion by investigating their intrinsic structural properties in both spatial and temporal dimensions. In this case, the underlying temporal correlations among adjacent silhouette frames are discovered by solving an adaptive time-series data alignment optimization problem using dynamic time warping (DTW). Furthermore, we build a part-based shape model to capture the spatial structural information on human silhouettes by sparseness constrained nonnegative matrix factorization (NMF)-based local feature learning, which is capable of well modeling the shape variation properties of human silhouettes. Experimental results on several challenging datasets demonstrate the effectiveness of our method.