Rotation & Viewpoint Angle Prediction in Capsule Network

Abstract

Convolutional neural network (CNN) is effective in detecting features and classifying object but less effective in exploring spatial relationships among the features. Capsule network introduces stacking layers called capsules and routing algorithm. Such a capsule structure is proved to handle spatial relationships better than CNN architecture. This paper aims at exploring capsule network ability to adapt with the rotation and viewpoint change in image recognition for MNIST, SmallNORB and CAS-PEAL datasets compared to CNN architecture (VGG-based network). The experimental results show that capsule network performs better than CNN for rotation estimation, whereas CNN architecture performs slightly better than capsule network for viewpoint change. The experiments also show that capsule network may have capability to generalize better in some untrained data for rotation and viewpoint change. Capsule network is quite promising architecture in classification and spatial context.