Section: New Results

Estimation of parametric motion models with deep neural networks

Participants : Juan Manuel Perez Rua, Patrick Bouthemy.

We have proposed an end-to-end learning architecture for estimating a parametric motion model for a moving scene. We handle motion outliers by using the supervised training trick that is used by stacked denoising auto-encoders. Here, we define motion outliers as regions of the image whose motion does not correspond with the estimated parametric motion model. In other words, we seek to find a parametrized dominant motion of the dynamic scene. We leverage stacked hourglass networks with a final hard-coded block corresponding to the global parametric motion model estimator. This block replaces the decoder part of a convolutional auto-encoder network, and it is end-to-end trainable since it involves linear operations only. Moreover, the hard-wired decoder allows the network to output the values of the parametric motion model given an input moving scene, even when the supervision acts on optical flow maps and not the motion model values. This means that our network is able to provide, as a by-product, a concise code that can be used as motion descriptor.

Collaborators: Tomas Crivelli and Patrick Pérez (Technicolor).