基于时序最大编码率约简的人体动作分割

Human Motion Segmentation (HMS) is an unsupervised learning task, aiming to partition videos into non-overlapping segments capturing different human motions.As an early processing step of human motion recognition and analysis, HMS is of great research importance and practical applicability.Recently, the self-expressive model based subspace clustering algorithms have emerged as dominant approaches in the HMS field.These models are based on the assumption that high-dimensional temporal data aligns with a Union-of-Subspaces (UoS) distribution andself-expressive models are proved to yield correct subspace clustering results under mild conditions.However, real-world temporal data may not align well with the UoS distribution, which hinders the model from achieving better performance.In this paper, we propose a novel HMS framework, named Temporal Rate Reduction Clustering (\trc{}), which is able to learn structured representations and segmentation result simultaneously.Specifically, \trc{} learns representations that align with a UoS distribution by incorporating the maximal coding rate reduction principle and ensures temporal consistency of representations by introducing the temporal Laplacian regularization.We conduct experiments and achieve state-of-the-art performances on four commonly used datasets of HMS, which validates the effectiveness of our framework.