We describe an unsupervised method to detect and segment portions of images
of live scenes that, at some point in time, are seen moving as a coherent
whole, which we refer to as objects. Our method first partitions the motion
field by minimizing the mutual information between segments.