Instructions For Accessing the Gist Algorithm Code

The code you want is saliency/src/INVT/ezvision.C.
To compile the code: make bin/ezvision
To run the code, in the saliency folder, run the command:

./bin/ezvision --in=input-image.ppm --out=display --K -ge-type=Std --save-gist=true --out=txt:prefix

The input (--in) can also be most other types, such as mpegs. The gist vector that we want to save in the code is in the form of Image gistVector.

the ge-type option specifies which algorithm to use here I selected the one presented in the PAMI paper, which is "Std"

--save-gist=true allows the gist features to be saved

--out=txt:prefix specifies prefix and the file type as text

the output should be prefix-Gist000000.txt

there should be 714 floats in there, which is 714 = 34 sub-channels (feature maps) x 21 values

* the first 16 subchannels are orientation channel
* the subsequent 12 are color channel, with the first six being Red/Green center-surround feature maps, the second six being Blue/Yellow center-surround maps.
* the last 6 sub channels are intensity channels, from dark/bright center-surround maps.

For each feature map there are 21 values that encompass average values of various spatial pyramids:
Value 0 is the average value of the entire feature map (this value is not used in the PAMI paper)
value 1 to 4 are the average value of each 2x2 quadrants of the feature map (this value is not used in the PAMI paper). with the spatial location being:
1 2
3 4

value 5 to 20 is the average value for each of the 4x4 grid of the feature map (this value IS used in the PAMI paper) with the spatial location being:
05 06 09 10
07 08 11 12
13 14 17 18
15 16 19 20

So if you would like to mimic the gist features in the PAMI exactly, just use the 5th to 20th features for each feature map.


Copyright © 2000-2009 by the University of Southern California, iLab and Prof. Laurent Itti