Computer Vision News - August 2016

18 Computer Vision News Research Research (a-b) Pair of images; (c) SED contour; (d) match positions ∈ DeepMatching; (e-f) geodesic distance from a pixel p (marked in blue) to all others (the brighter, the closer) (g-h) 100 nearest matches, using geodesic distance from the blue pixel. The approximate geodesic distance Input: (1) corresponding points; (2) Image contours (Fig. 3b); (3) two pixels p, p’ (Fig. 3a) Output: The approximate geodesic distance between p, p’ Method: 1. Compute the Voronoi partition (L) of the corresponding points ( ) (Fig. 3c) 2. Construct a graph G - whose nodes are the corresponding points and the edge weight is the geodesic distance between points 3. Calculate the geodesic distance from p and p ′ to their closest corresponding points and ′ (i.e. geodesic distance is restricted to the Voronoi cells of p and p’) 4. Calculate the shortest distance between and ′ using Dijkstra’s algorithm on G (Fig. 3d) Fig. 3 Fig. 2