Computer Vision News - July 2018

that are very far away, but actually they do not necessarily move. A tree which is close to us is displacing a lot, and a tree which is far away is moving just a little, but none of them actually move. Only the person or the animal is moving. There are a lot of connected topics like object segmentation, depth estimation, and motion segmentation which we have to consider all at the same time. Pia explains: “ We look at how the camera is moving. The camera can be translating, the camera can be rotating, and an object can be moving. All those three ingredients are part of the motion in the world which creates the optical flow. We first estimate the camera rotation and we subtract this off the optical flow, such that we only have a translational flow field, which has the camera translation and the object motion. If you just look on the direction of the optical flow, not on the magnitude, then you can see which objects are moving into a different direction. We developed a system which is considering the anti- optical flow and the derived flow likelihood, which computes how likely is a motion model given the observed optical flow .” In terms of next steps, Pia tells us that she will focus more on solving all those problems at the same time – depth estimation, motion segmentation, object recognition, as well as estimating optical flow . She predicts that if you solve all those problems together, they can support each other. Pia Bideau presented her work on a poster at CVPR on June 19. “If you solve all those problems together, they can support each other!” Tuesday 11 Pia Bideau