Dense Optical Flow using RAFT

Abstract

RAFT is a deep network architecture for the detection of optical flow in the images. The RAFT model relates the per pixel motion between images even for minor changes in the position of the objects. It also updates the flow of field through recurrent units that perform lookups on the performance of the model. RAFT also works well with different datatypes and also it has better efficiency, training speed and count of parameters. Experiments were performed by using different parameters and also by changing certain values in the model itself. One cycle learning was also used to find the best parameters for the model. We also found that the RAFT model performs better than most of the other existing models for optical flow calculation in to images.