Optical Flow


Large motions remain a challenge for current optical flow algorithms. Traditionally, large motions are addressed using multi-resolution representations like Gaussian pyramids. To deal with large displacements, many pyramid levels are needed and, if the object is small, it may not be visible at the highest pyramid levels. To address this we describe images using a channel representation (CR) and replace the standard brightness constancy assumption with a new descriptor constancy assumption, where the descriptor is the channel representation element at each pixel. Loosely, CRs can be thought of as over-segmenting the scene into layers based on some image feature and representing an image as a collection of local distributions, one at each pixel. If the appearance of a foreground object differs from the background then its descriptor will be different and it will be represented in different layers. We can thus smooth the layers of the CR without mixing together the foreground and background. Thus, we can apply pyramid techniques to the CR to capture large motions without completely losing small objects or blurring across motion boundaries. On the MPI-Sintel benchmark, we are more accurate than the baseline for fast motions and near motion boundaries.


Graduate Students



  • Laura Sevilla-Lara, Deqing Sun, Erik G. Learned-Miller, Michael J. Black.
    Optical Flow Estimation with Channel Constancy
    European Conference on Computer Vision (ECCV), 2014. (To appear)