The great convergence is upon us, here is clue #734: Andrew Davison mentioning recent work in optical flow using CNNs.
Whoa, this is a wake up call... CNN based learned optical flow (trained on synthetic flying chairs!) running at 10fps on a laptop which claims state of the art accuracy among real-time optical flow methods. So time for those of us working on non learning-based vision to pack up and go home?
This is a pretty powerful statement from one of the specialist of SLAM. Here is the paper:
FlowNet: Learning Optical Flow with Convolutional Networks by Philipp Fischer, Alexey Dosovitskiy, Eddy Ilg, Philip Häusser, Caner Hazırbaş, Vladimir Golkov, Patrick van der Smagt, Daniel Cremers, Thomas Brox
Convolutional neural networks (CNNs) have recently been very successful in a variety of computer vision tasks, especially on those linked to recognition. Optical flow estimation has not been among the tasks where CNNs were successful. In this paper we construct appropriate CNNs which are capable of solving the optical flow estimation problem as a supervised learning task. We propose and compare two architectures: a generic architecture and another one including a layer that correlates feature vectors at different image locations.
Since existing ground truth data sets are not sufficiently large to train a CNN, we generate a synthetic Flying Chairs dataset. We show that networks trained on this unrealistic data still generalize very well to existing datasets such as Sintel and KITTI, achieving competitive accuracy at frame rates of 5 to 10 fps.
Liked this entry ? subscribe to Nuit Blanche's feed, there's more where that came from. You can also subscribe to Nuit Blanche by Email, explore the Big Picture in Compressive Sensing or the Matrix Factorization Jungle and join the conversations on compressive sensing, advanced matrix factorization and calibration issues on Linkedin.