gbvs.rar_gbvs_saliency map_显著图_视觉显著图

Graph-Based Visual Saliency (MATLAB source code) http://www.klab.caltech.edu/~harel/share/gbvs.php by Jonathan Harel jonharel@gmail.com California Institute of Technology ================================================================================ This is an installation and help file for the saliency map (MATLAB) code here. What you can do with this code: 1) Compute a "Graph-Based Visual Saliency" map for an image or image sequence (as described in J. Harel, C. Koch, and P. Perona. "Graph-Based Visual Saliency", NIPS 2006 http://www.klab.caltech.edu/~harel/pubs/gbvs_nips.pdf) 2) OR -- also -- compute the standard Itti, Koch, Niebur (PAMI 1998) saliency map. .. or combinations thereof (see below). ================================================================================ Step-by-step start-up procedure: (1) Add gbvs to your path: Change into the directory containing this file, and run in matlab: matlab> gbvs_install If you are on a shared machine, you may get an error message such as: Warning: Unable to save path to file '/opt/matlab/toolbox/local/pathdef.m' In savepath at 162 In gbvs_install at 5 In that case, comment out the savepath (i.e., 'savepath' => '% savepath') command in gbvs_install.m, and add this line to your startup.m file: run ???/gbvs_install where "???" is replaced by the main gbvs/ directory, which contains the gbvs_install function (2) **Unless you downloaded Windows binaries**, you will need to compile: matlab> gbvs_compile Notes: (a) If this is your first time compiling mex files in Windows, you may have to run matlab> mex -setup and follow the instructions (typically, enter a number, to select a co- mpiler. then you can run "gbvs_compile"; if it doesn't work, run "mex -setup" again to select a different compiler, run "gbvs_compile" again, etc.) (b) Ignore these messages if they appear: rm: */*.mex*: No such file or directory 'rm' is not recognized as an internal or external command, operable program or batch file. They are harmless. (3) Now you are ready to compute GBVS maps: Demonstrations: matlab> simplest_demonstration see demo/demonstration.m for more complicated demo or run: matlab> demonstration Basic Usage Example: matlab> out = gbvs( 'samplepics/1.jpg' ); Now, out.master_map contains your saliency map, and out.master_map_resized is this saliency map interpolated (bicubic) to the resolution of the original image. For video (not static images): You need to pass into gbvs() previous frame information, which is returned on output at every call to gbvs(). See demo/flicker_motion_demo.m Here is the heart of it: motinfo = []; % previous frame information, initialized to empty for i = 1 : N [out{i} motinfo] = gbvs( fname{i}, param , motinfo ); end ================================================================================ Helpful Notes: (A) inputs of gbvs(): * the first argument to gbvs() can be an image name or image array * there is an optional, second, parameters argument (B) outputs of gbvs(): * all put into a single structure with various descriptive fields. * the GBVS map: master_map (interpolated to the resolution of the input image: master_map_resized) * master saliency map for each channel: feat_maps (and their names, map_types) * all intermediate maps to create the previous two (intermed_maps). see gbvs.m for details (C) the parameter argument: * initialized by algsrc/makeGBVSParams.m -- read that for details Some sparse notes on fields of the parameter argument: the two 'sigma' parameters control the extent to which information can travel across the image: (1) sigma_frac_act sigma parameter in activation step of GBVS (fraction of image width) (2) sigma_frac_norm sigma parameter in normalizaiton step of GBVS (fraction of image width) cyclic_type default is 1. use 2 to have edge weights computed with non-cyclic distance rules. this gives rise to a center bias. tol tolerance parameter. governs how accurately the princi- pal eigenvector calculation is performed. change it to higher values to make things run faster. levels the resolution of the feature maps used to compute the final master map, relative to the original image size multilevels determines whether to instantiate one-level or multi- level lattice of graph nodes for each feature map useIttiKochInsteadOfGBVS -- set this to 1 to compute Itti/Koch map (D) Notes on feature maps: * are produced by util/getFeatureMaps.m * by default, color, intensity, orientation maps are computed. which channels are used is controlled by the parameters argument. in part- icular, you can choose which of these is included by editing the params.channels string (see algsrc/makeGBVSParams.m). you can set their relative weighting also in the parameters. If you want to introduce a new feature channel, put a new function into util/featureChannels/ . Make sure to edit the channels string appropria- tely. Follow pattern of other channels for proper implementation. (E) If you want to compare saliency maps to fixations (e.g., inferred from scanpaths recorded by an eye-tracker), use: matlab> score = rocScoreSaliencyVsFixations(salmap,X,Y,origimgsize) This outputs ROC Area-Under-Curve Score between a saliency map and fixat- ions. salmap : a saliency map X : vector of X locations of fixations in original image Y : vector of Y locations of fixations in original image origimgsize : size of original image (should have same aspect ratio as saliency map) ================================================================================ Copyright Notices for borrowed code: saltoolbox/ directory -- adapted from: Dirk Walther, http://www.saliencytoolbox.net/ ================================================================================ Revision History first authored 8/31/2006 Revised 4/25/2008 Revised 6/5/2008 Revised 6/26/2008 added Itti/Koch algorithm Revised 8/25/2008 added Flicker/Motion channels