The human ability to detect and segment moving objects works in the presenceof multiple objects, complex background geometry, motion of the observer, andeven camouflage. In addition to all of this, the ability to detect motion isnearly instantaneous. While there has been much recent progress in motionsegmentation, it still appears we are far from human capabilities. In thiswork, we derive from first principles a new likelihood function for assessingthe probability of an optical flow vector given the 3D motion direction of anobject. This likelihood uses a novel combination of the angle and magnitude ofthe optical flow to maximize the information about the true motions of objects.Using this new likelihood and several innovations in initialization, we developa motion segmentation algorithm that beats current state-of-the-art methods bya large margin. We compare to five state-of-the-art methods on two establishedbenchmarks, and a third new data set of camouflaged animals, which we introduceto push motion segmentation to the next level.
展开▼
