Learning Self-Expressive Representations via Nearest Subspace Model for Motion Segmentation

Subspace clustering methods based on self-expressive model aim to represent each data point as a linear combination of data points belonging to the same subspace. However, conventional methods based on a self-expressive model are known to suffer from a connectivity problem in which data points belonging to the same subspace are not sufficiently connected. To address this problem, we propose a subspace clustering method by learning self-representations using the nearest subspace model. The adoption of the nearest subspace model leads to the dense connection between data lying on the same subspace. In addition, we incorporate the update of the subspace model into the optimization process to refine the model. Evaluation experiments were conducted using real-world data, showing the effectiveness of clustering in the self-expressive model based on the nearest subspace model.