FlowCaps: Optical Flow Estimation with Capsule Networks For Action Recognition
This addresses optical flow estimation for action recognition, potentially improving efficiency and interpretability, but appears incremental as it adapts capsule networks to an existing task.
The paper tackles optical flow estimation by proposing FlowCaps, a capsule network-based architecture, which aims to improve correspondence matching, generalization, data efficiency, and computational complexity compared to CNN methods, though no concrete performance numbers are provided.
Capsule networks (CapsNets) have recently shown promise to excel in most computer vision tasks, especially pertaining to scene understanding. In this paper, we explore CapsNet's capabilities in optical flow estimation, a task at which convolutional neural networks (CNNs) have already outperformed other approaches. We propose a CapsNet-based architecture, termed FlowCaps, which attempts to a) achieve better correspondence matching via finer-grained, motion-specific, and more-interpretable encoding crucial for optical flow estimation, b) perform better-generalizable optical flow estimation, c) utilize lesser ground truth data, and d) significantly reduce the computational complexity in achieving good performance, in comparison to its CNN-counterparts.