Conventional methods for optical flow estimation are inappropriate for applications that demand real-time operation and small memory capacity. Deep learning based networks, such as FlowNet[l] and FlowNet2[2], have high accuracy and fast running time, ...
Conventional methods for optical flow estimation are inappropriate for applications that demand real-time operation and small memory capacity. Deep learning based networks, such as FlowNet[l] and FlowNet2[2], have high accuracy and fast running time, but require more than 160M of parameters. Therefore, we introduce tow deep learning based networks called PWC-Net[3] and LiteFlowNet[4] that have compact number of parameters and also have effective performances. Introducing two network models of compact but effective CNN models for optical flows, called PWC-Net and LiteFlowNet.