Computer Vision News - July 2016

 Line1: Initialization step, where k initial cluster centers in the CIELAB color space and location x,y. The cluster center are sample along equally spaced a grid with S=√(imageSize/k )  Line 2: Initialization all labels l(i) to -1 (i.e. no label) and the distance d(i) to infinity.  Line 5-10: The assignment step, each pixel i is associated with the nearest cluster center in a bounded region 2S ⋅ 2S. This is the speeding up step, because limiting the size of the search region significantly reduces the number of distance calculations.  Line 14: The process ends when the residual error E is below some fixed threshold The distance D is defined as follows: The distance D normalizes color proximity and spatial proximity by their respective maximum distances, where: • d c ,d s are the color and spatial normalization factor • m is the compactness term. m allows us to weigh the relative importance between color similarity and spatial proximity. When m is large, spatial proximity is more important and the resulting SuperPixels are more compact. Examples: Now we will see how this works with some examples. Let’s say you want to segment the following image into 4 regions: Pony body, Pony hear, the Tuba and the background. Some (and there are many others) popular segmentation algorithms for doing so could be the DBSCAN or the spectral clustering. Let’s consider the spectral clustering: in order to work with this algorithm you have to construct an affinity matrix between each pair of pixels in the image. The dimension of the above image is 348 x 623 = 216804 pixels. This means that the affinity matrix is of the size 216804x216804 which is quite big. We will reduce the number of pixels to somewhere between 10 to 20% by using the SLIC SuperPixels algorithm. A reduction of 15% makes about 150,000 SuperPixels . This is how it looks: Computer Vision News Tool 21 = 2 + ( / ) 2 ⋅ 2 Tool