Section: New Results

Motion saliency in video sequences

Participants : Léo Maczyta, Patrick Bouthemy.

Dynamic (or motion) saliency is a means to detect unexpected or rare dynamic behaviors in video sequences acquired by a stationary or a mobile imaging device. Finding salient dynamic information in each image of a sequence is indeed crucial in many situations. We aim to extract saliency only from motion information, and to exhibit salient motion in contrast to its space-time context with no prior on the nature of both. So far, we have investigated a simpler problem than saliency map estimation. We deal with the classification of each image of a sequence as dynamically salient or not, that is, containing salient motion or not. We have explored convolutional neural network (CNN). We have designed two different networks. The first one relies on two intensity images, the first input image and the second image warped with the parametric dominant motion estimated between the two input images. The second one takes as input the difference between the computed optical flow and parametric dominant flow.

Collaborators: Olivier Lemeur (EPC Sirocco, Inria Rennes - Bretagne Atlantique).