Unsupervised Optical Flow Estimation with temporal smoothing

Advisor: Prof. Amir Barati Farimani, Mechanical Engineering, Carnegie Mellon University

System Overview:

• Designed an unsupervised version of the Flownet-C architecture for optical flow estimation
• Formulated a temporal smoothing loss term which penalizes large changes in consecutive optical flow maps
• Generated temporally smoother optical flow maps producing more temporally consistent warped images.

Results:

Image Frame	Optical Flow Map

Loss curves comparison on test dataset:

• The project report is available here.