Computer Vision News - July 2016

The SLIC (Simple Linear Iterative Clustering) SuperPixels algorithm by Achanta et al . is celebrating 6 years and this is the first time that the image-processing Matlab toolbox (ver 2016a, released in March) was shipped with a built-in implementation of this algorithm. It is therefore a great opportunity to write about this useful algorithm on Computer Vision News and show some applications of it. What is SuperPixels about? Pixels are the common unit of many computer vision algorithms. Pixels, however, are a consequence of the image discrete representation and aren’t the natural image representation. SuperPixels algorithms group pixels into perceptually meaningful regions, which can be used to replace the pixel-based representation of an image. SuperPixels has many desired properties, amongst which are: (1) images size reduction from thousands of pixels to only a few hundred; (2) conservation of the natural image structure ( i.e. uniform in color and texture); (3) little information loss; (4) a convenient primitive from which to compute local image features while reducing the complexity. The algorithm itself: The beauty of this algorithm is that it is straightforward, easy to understand and useful in many computer vision application. In a nutshell, the SLIC SuperPixels adapts the k-means clustering algorithm to efficiently generate SuperPixels of nearly uniform size whose boundaries closely match the true image boundaries. This representation can be thought of over segmentation of the original image. Here is the pseudocode of the algorithm, with more details following: 20 Computer Vision News Tool SLIC SuperPixels SLIC SuperPixel algo. 1. Initialized cluster center = [ , , , , ] sampling pixels at regular grid steps 2. For each pixel, set label ( ( ) = −1) and the distance ( ) = ∞ 3. Loop 4. For (each cluster center ) 5. For (each pixel in 2 ⋅ 2 region around ) 6. Computer the distance between and 7. If ( <d(i)) 8. d(i) = 9. l(i) = k 10. end if 11. end for 12. end for 13. Update cluster centers 14. Compute residual error E 15. Until E < threshold Tool