Optical Flow Prediction in Auto Driving from Single Image via Conditional Variational Auto-Encoder

In auto-driving tasks, visual prediction is very meaningful and difficult, because the motion of the predicted target has many possible outputs. It is very effective to predict the motion of the target by optical flow, but the output of the previous optical flow models is fixed. In this paper, we proposed an optical flow prediction model via Conditional Variational Auto-Encoder. In experiment, the model can effectively predict a variety of possible optical motions in the real world. A number of experimental results show that our models outperform all prior state-of-the-art on the test of a recent optical flow prediction competition.