Optical Flow Estimation Using a Non-Local Convolutional Network

Convolutional neural network(CNN) models for optical flow estimation based on coarse-to-fine method are usually difficult to obtain accurate estimates of large displacement motions in the rough layer, so that the estimation error will be passed to the final estimation result. This article proposes an effective convolutional neural network model for optical flow estimation called NTFlow. NTFlow uses a non-local convolutional layer to obtain the correlation of the full feature map, and constrains the estimate of the larger error in the loss function. Experiment results show that our network can get accurate estimation results on public data sets, and the proposed loss function is very robust.