function imdb = updateImdbVOC()

baseDir = '/home/bshuai/datasets/VOC2012/';

imdb = load(fullfile(baseDir, 'imdb-train-val-update-disk.mat'));

nImgs = numel(imdb.images.name);

for i = 1: nImgs

imdb.images.data{i} = sprintf(imdb.paths.image, imdb.images.name{i});

imdb.images.labels{i} = sprintf(imdb.paths.classSegmentation, imdb.images.name{i});

trainIdx = imdb.images.set == 1 & imdb.images.segmentation ;

nTrainImgs = sum(trainIdx);

train_data = imdb.images.data(trainIdx);

train_label = imdb.images.labels(trainIdx);

tid = ticStatus('Calculating mean Image...',1,1);

rgbMean = zeros(1,1,3);

frequency = zeros(1,22);

nPixels = 0;

for i = 1 : nTrainImgs

I = single(imread(train_data{i}));

[h,w,~] = size(I);

nPixels = nPixels + h*w;

rgbMean = rgbMean + (sum(sum(I,1),2));

label = imread(train_label{i});

% transform label

label = mod(label + 1, 255);

frequency = frequency + hist(label(:), 0:21);

tocStatus(tid,i/nTrainImgs);

rgbMean = rgbMean / nPixels;

imdb.rgbMean = rgbMean;

imdb.classFrequency = frequency;

imdb.meta.sets = {'train', 'val', 'test'};

classes = imdb.classes;

classFrequency = imdb.classFrequency;

images = imdb.images;

paths = imdb.paths;

rgbMean = imdb.rgbMean;

sets = imdb.sets;

meta = imdb.meta;

save(fullfile(baseDir, 'imdb-train-val-update-disk.mat'), ...

'classes', 'classFrequency', 'images', 'paths', 'rgbMean', ...

'sets', 'meta');