File size: 4,898 Bytes
d4035c1 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 |
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
SOURCE README FOR AUTOMATIC PHOTO POPUP AND GEOMETRIC CONTEXT
Derek Hoiem (dhoiem@cs.cmu.edu)
01/08/2010
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
LICENSE
Copyright (C) 2007 Derek Hoiem, Carnegie Mellon University
This software is available for only non-commercial use. See the attached
license in LICENSE.txt.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
REFERENCES
This code implements the following papers. Note that the implementation may not be exact.
Please cite one or more of these papers, depending on the use.
D. Hoiem, A.A. Efros, and M. Hebert, "Geometric Context from a Single Image", ICCV 2005.
D. Hoiem, A.A. Efros, and M. Hebert, "Recovering Surface Layout from an Image", IJCV, Vol. 75, No. 1,
October 2007.
D. Hoiem, A.A. Efros, and M. Hebert, "Automatic Photo Pop-up", ACM SIGGRAPH 2005.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
Note about VERSIONS
This contains two versions of the "Automatic Photo Pop-up" and "Geometric
"Context" code. The first version is from the SIGGRAPH 2005 / ICCV 2005 papers
and involves the functions photoPopup and APPtestImage. The second version
is from the IJCV 2007 paper and involves the functions photoPopupIjcv and
ijcvTestImage. Note that these implementations may not be exact. In particular,
some changes to features were made to improve speed, which may have small effects
on accuracy. Further, the multiple segmentation algorithm is random, so different
runs will not produce identical results.
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
How to RUN:
First, appropriately replace the path for the segment command in im2superpixels.m.
Geometric Context version 1 (ICCV 2005):
[labels, conf_map, maps, pmaps] = APPtestImage(image, imsegs, vClassifier, hClassifier, segDensity)
image should be a double color image
imsegs is the superpixel structure (if empty, will be computed)
The remaining three arguments are stored in data/classifiers_08_22_2005.mat
[labels, conf_map, imsegs] = ...
APPtestDirectory(segDensity, vClassifier, hClassifier, imdir, imfn, varargin)
Processes a directory of images with given filenames. Optional last argument is where to
store displayed results.
Geometric Context version 2 (IJCV 2007):
[pg, data, imsegs] = ijcvTestImage(im, imsegs, classifiers, smaps, spdata, adjlist, edata);
The only required fields are im and classifiers.
Example usage:
load '../data/ijcvClassifier.mat'
classifiers=load('../data/ijcvClassifier_indoor.mat');
[pg, data, imsegs] = ijcvTestImage(im, [], classifiers);
[cimages, cnames] = pg2confidenceImages(imsegs, {pg});
% only keep ground pourous sky vertical [1 5 7 8]
geom_c_map = cimages{1}(:,:,[1 5 7 8]);
Classifiers trained on indoor and outdoor images are provided in the data directory.
[pg, smaps, imsegs] = ijcvTestImageList(fn, imsegs, classifiers, laboutdir, confoutdir)
Processes all images whose names are given in fn. As above, imsegs can be empty.
Automatic Photo Pop-up (SIGGRAPH 2005, IJCV 2007)
photoPopup(fnData, fnImage, extSuperpixelImage, outdir)
This is the 2005 version.
fnData is the filename containing the classifiers (data/classifiers_08_22_2005.mat).
If extSuperpixelImage is empty, it will be computed.
photoPopupIjcv(fnData, fnImage, extSuperpixelImage, outdir)
This is the IJCV 2007 version. See notes above.
Training:
For the IJCV version, see src/ijcv06/ijcvMultiSegScript.m
This function performs both training and evaluation. It requires some edits. It is currently
setup for cross-validation, but it should be straightforward to use separate train and test sets.
The Geometric Context Dataset (separate download) contains rand_indices which specifies the
train/test splits.
Other useful functions:
src/util/splitpg.m
src/util/pg2confidenceImages.m
src/util/pg2*
src/util/write*
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%%
NOTES:
Superpixels:
I use the segmentation code provided by Felzenszwalb and Huttenlocher
at people.cs.uchicago.edu/~pff/segment/ to create the superpixels in my
experiments. The first three arguments (sigma, k, min) that I use are
0.8 100 100. I've included a pl script for segmenting a directory
that you may find useful. You can also use a different program to create
the superpixel image. That image should have a different RGB color for
each segment without drawn boundaries between segments.
(C) Derek Hoiem, Carnegie Mellon University, 2007
|