About

This version of turbopixels code is almost exactly same as the original implementation from Alex Levinshtein based on [1]. All the minor modifications in the code were mostly done to get the code to compile on 64-bit linux (ubuntu 18.04 with Matlab 2019b). I haven't tested it on any other systems.

Part of disclaimer from original readme:

The affinity function that is used in the code is simple and is not sufficient to provide good results on all possible images. The domain of the images affects the choice of the affinity function.

Installation:

• clone this repository and move in the root directory

$ cd turbopixel

• The subdirectory 'lsmlib' needs to be added to matlab's path. This is done in demo_superpixels.m, or you can do it manually by typing the following command into matlab command window

>>> addpath('lsmlib');

• Run make to compile c implementations. To work, MEX needs to be properly configured. Ignore the warnings.

>>> make;

• Run demo_superpixels.m to run turbopixels on sample lizard.jpg image provided.

>>> demo_superpixels;

Examples:

In additions to demo_superpixels following superpixels may also be run as follows

Getting 1500 superpixels without intermediate display

img = im2double(imread('lizard.jpg'));
[phi,boundary,disp_img] = superpixels(img, 1500);
imshow(disp_img);

Getting superpixels with an intermediate display every 5 timesteps

img = im2double(imread('lizard.jpg'));
[phi,boundary,disp_img] = superpixels(img, 1500, 5);
imshow(disp_img);

References

[1] Levinshtein, Alex, et al. "Turbopixels: Fast superpixels using geometric flows." IEEE transactions on pattern analysis and machine intelligence 31.12 (2009): 2290-2297.