Gist描述子

场景识别描述子Gist代码

 

Gist描述子主要用于场景识别，是由MIT的Antonio Torralba提出的。文章发表在IJCV，题目：Modeling the shape of the scene: a holistic representation of the spatial envelope

GIST Descriptor (Matlab code)

Download:

Download all the matlab code and examples here: gistdescriptor.zip

Computing the gist descriptor:

To compute the gist descriptor on an image use the function LMgist. The next example reads one image and computes the descriptor (the images demo1.jpg and demo2.jpg are available inside the gistdescriptor.zip file).

% Load image
img = imread('demo2.jpg');

% GIST Parameters:
clear param
param.orientationsPerScale = [8 8 8 8]; % number of orientations per scale (from HF to LF)
param.numberBlocks = 4;
param.fc_prefilt = 4;

% Computing gist:
[gist, param] = LMgist(img, '', param);

Visualization:

To visualize the gist descriptor use the function showGist.m. Here there is an example of how to use it:

% Visualization
figure
subplot(121)
imshow(img)
title('Input image')
subplot(122)
showGist(gist, param)
title('Descriptor')

Image similarities:

When computing image similarities, it might be important to normalize the image size before computing the GIST descriptor. This can be achieved by setting the image size inside the param struct (using the field param.imageSize). The LMgist function will resize and crop each image to match the specified size before computing the gist descriptor. The resizing operation will not affect the aspect ratio of the original image. The crop will be centered and the image will be resize so that the cropped region preserves as much as possible from the original input image. Here is an example:

% Load images
img1 = imread('demo1.jpg');
img2 = imread('demo2.jpg');

% GIST Parameters:
clear param
param.imageSize = [256 256]; % it works also with non-square images (use the most common aspect ratio in your set)
param.orientationsPerScale = [8 8 8 8]; % number of orientations per scale
param.numberBlocks = 4;
param.fc_prefilt = 4;

% Computing gist:
gist1 = LMgist(img1, '', param);
gist2 = LMgist(img2, '', param);

% Distance between the two images:
D = sum((gist1-gist2).^2)

Image collections:

The first call to LMgist will precompute the filters in the frequency domain and store them in param.G, subsequent calls will be faster.

% GIST Parameters:
clear param
param.imageSize = [256 256]; % set a normalized image size
param.orientationsPerScale = [8 8 8 8]; % number of orientations per scale (from HF to LF)
param.numberBlocks = 4;
param.fc_prefilt = 4;

% Pre-allocate gist:
Nfeatures = sum(param.orientationsPerScale)*param.numberBlocks^2;
gist = zeros([Nimages Nfeatures]); 

% Load first image and compute gist:
img = imread(file{1});
[gist(1, :), param] = LMgist(img, '', param); % first call
% Loop:
for i = 2:Nimages
   img = imread(file{i});
   gist(i, :) = LMgist(img, '', param); % the next calls will be faster
end

The script demoGist.m shows a few more examples and also how it works with non-square images. The function LMgist can also work the LabelMe toolbox.

http://people.csail.mit.edu/torralba/code/spatialenvelope/

project page：