Sparse RecoveryAlgorithmResults Original signal x = x k + u, where x k has k large coefficients and u is noise. Acquire measurements Ax = y. If |x|=n, A is an m x n matrix, and usually m = O(k log n) << n. Recover vector x* from y. The usual guarantee is ||x* – x|| 1 < C ||u|| = C min k-sparse x’ ||x' – x|| 1, for some constant C. In other words, x* is close to the best possible k-sparse approximation of x. High Level Overview Use a distance preserving embedding to map the problem under EMD to a problem under the L 1 norm. Solve as a standard sparse recovery problem in the L 1 norm. The Pyramid Mapping Use a distance preserving embedding to map the problem under EMD to a problem under the L 1 norm. For each j = 0, 1, 2,..., form a grid with cell length d j = 2 j, and a mapping P j, where each entry of P j x corresponds to the total mass of a cell of the grid. The mapping Px is the concatenation [d 0 P 0 x, d 1 P 1 x, d 2 P 2 x,...]. If the grids are shifted by a random vector, E[||Px 1 – Px 2 || 1 ] < O(log n) ||x 1 – x 2 || EMD, and ||Px 1 – Px 2 || 1 > ||x 1 – x 2 || EMD for images x 1 and x 2 [3]. Hence P is a linear, distance preserving embedding from the L 1 norm to the EMD norm. Also, if x is k-sparse, Px is k log(n)-sparse. Full Algorithm Let A be any matrix enabling k log(n)-sparse recovery, eg. from [1]. Measurements y = APx, where P is the pyramid mapping from above. Recovery: Compute “A -1 y” using a sparse recovery algorithm, and then compute x* = P -1 A -1 y using an “inverse” pyramid mapping. In our experiments, we use the following constraints during the sparse recovery portion of the algorithm to reduce the number of measurements |y|: x is non-negative for real life images. The large coefficients of Px form a “tree”. Theoretical We find a set of m x n matrices B, m = O(k log n/k), such that for any x, given Bx, we can compute x* such that ||x* - x|| EMD < C min k-sparse x’ ||x' – x|| EMD, with high probability, in O(k log n) time, for a constant C. Experimental We use Sequential Sparse Matching Pursuit (SSMP) [1] for the Sparse Recovery portion of the algorithm. We test our algorithm on the images such as the one below to the left. After recovery, we attempt to locate five star-like objects in x*. Each point on the graph on the right is the median over 15 trials of the average distance from recovered stars to the actual stars. If fewer than 5 stars are found, the distance is assumed to be infinite and is not displayed. The EMD algorithm recovers the stars using substantially fewer measurements than standard SSMP. Earth Movers Distance (EMD) x represents a 2 dimensional image. If x 1 (green) and x 2 (purple) are binary images such that ||x 1 || 1 = ||x 2 || 1, then ||x 1 – x 2 || EMD is the minimum cost matching shown to the right. There are natural generalizations to non-binary vectors with unequal L 1 mass. Problem Statement We want to perform sparse recovery as above, but we would like to recover x* such that x* is close to x under EMD rather than L 1. EMD corresponds well with our perceptual notion of image similarity, and is used in computer vision algorithms. By contrast, even small translations in an image result in almost maximal L 1 difference [4]. References [1]Berinde, Indyk. Sequential sparse matching pursuit. Allerton [2]Gupta, Indyk, Price. Sparse recovery for earth mover distance. Allerton [3]Indyk, Thaper. Fast image retrieval via embeddings. ICCV [4]Rubner, Tomasi, Guibas. The earth mover's distance as a metric for image retrieval. IJCV Sparse Recovery for Earth Mover Distance MADALGO – Center for Massive Data Algorithmics, a Center of the Danish National Research Foundation Eric Price MIT Piotr Indyk MIT Rishi Gupta MIT