Grassmannian learning mutual subspace method for image set recognition

This paper addresses the problem of object recognition given a set images as input (e.g., multiple camera sources and video frames). Convolutional neural network (CNN)-based frameworks do not exploit these sets effectively, processing pattern observed, capturing underlying feature distribution it does consider variance in set. To address this issue, we propose Grassmannian learning mutual subspace method (G-LMSM), NN layer embedded on top CNNs that can process image more effectively be trained an end-to-end manner. The is first represented by low-dimensional then matched with dictionary subspaces similarity their canonical angles, interpretable easy to compute metric. key idea G-LMSM are learned points Grassmann manifold, optimized Riemannian stochastic gradient descent. stable, efficient theoretically well-grounded. We demonstrate effectiveness our proposed hand shape recognition, face identification, facial emotion recognition.