Computer Vision News - August 2016

Computer Vision News Computer Vision News Research 17 Research Method The method consists of two main steps (Fig. 1 above): first it calculates edge- aware dense interpolation instead of the classical sparse-to-dense interpolation. To do that, it obtains a set of corresponding points for the two images using DeepMatching and contours of the first image using Structured Edge Detector (SED) . Next, for each corresponding point in p m ∈ M it calculates the set of k closest points, then marks those points as N k (p m ) (Fig. 2). The closest points are calculated using the approximate geodesic distance (Fig. 3). The method is summarized in the boxes below. EpicFlow Input: ● Pair of images: , ′ Output: ● Th e Optical flow e stimation between the two images Method (1) Edge aware dense interpolation field (a) Obtain set of corresponding points for the two images using DeepMatching (b) Obtain contours of the first image using Structured Edge Detector (SED) (c) For each point ∈ calculate the set of k closest points using the approximate geodesic distance (see below and see also Fig. 2) (d) Compute the (dense) correspondence field for each ∈ (e) For each pixel in (i) Set ( ) = ( ( )) - ( ) (2) Variational refinement (a) Dense matching interpolation (b) Non-local smoothness term (c) Classical solver (fixed point iterations, successive over relaxation) Geodesic Voronoi diagram . is the clustering partition defined by the Voronoi diagram (Fig. 3c) and ( ) assigns a pixel to its closest match according to the geodesic distance, i.e. we have ( ) = ( , ) . Fig. 1