Measure

Geometric Vision and Learning

Learning methods for multiframe Optical Flow Estimation

Work by Pierre Godet, who defended its PhD on friday, 22th january 2021, in collaboration with Alexandre Boulc’h at Valeo.ai and Aurélien Plyer ONERA.

Pierre first derived a multiframe Lucas-Kanade method based on temporal models identified by PCA for PIV application. Turning towards deep learning approaches, he proposed “STaRFlow” a multiframe optical flow estimation relying on a spatio-temporal recurrent cell with occultation handling. STaRFlow is among state-of-the-art methods on Kitti and Sintel benchmark and shows remarkable generalization capabilities to videos in real contexts.
[PDF] [GitHub] [Video]