function predict_depth()

% setup vlfeat

run( '../libs/vlfeat-0.9.18/toolbox/vl_setup');

% setup matconvnet

dir_matConvNet='../libs/matconvnet/matlab/';

addpath(genpath(dir_matConvNet));

run([dir_matConvNet 'vl_setupnn.m']);

opts=[];

opts.useGpu=true;

opts.inpaint = true;

opts.normalize_depth = false; % limit depth to [0,1]

opts.imdir = '/path/to/image/dir';

opts.out_h5 = '/path/to/save/output/depth.h5';

% these should point to the pre-trained models from:

% https://bitbucket.org/fayao/dcnf-fcsp/

opts.model_file.indoor = '../model_trained/model_dcnf-fcsp_NYUD2';

opts.model_file.outdoor = '../model_trained/model_dcnf-fcsp_Make3D';

fprintf('\nloading trained model...\n\n');

mdl = load(opts.model_file.indoor);

model.indoor = mdl.data_obj;

mdl = load(opts.model_file.outdoor);

model.outdoor = mdl.data_obj;

if gpuDeviceCount==0

fprintf(' ** No GPU found. Using CPU...\n');

opts.useGpu=false;

end

imnames = dir(fullfile(opts.imdir),'*');

imnames = {imnames.name};

N = numel(imnames);

for i = 1:N

fprintf('%d of %d\n',i,N);

imname = imnames{i};

imtype = 'outdoor';

img = read_img_rgb(fullfile(opts.imdir,imname));

if strcmp(imtype, 'outdoor')

opts.sp_size=16;

opts.max_edge=600;

elseif strcmp(imtype, 'indoor')

opts.sp_size=20;

opts.max_edge=640;

end

depth = get_depth(img,model.(imtype),opts);

save_depth(imname,depth,opts);

end

function save_depth(imname,depth,opts)

dset_name = ['/',imname];

h5create(opts.out_h5, dset_name, size(depth), 'Datatype', 'single');

h5write(opts.out_h5, dset_name, depth);

function depth = get_depth(im_rgb,model,opts)

% limit the maximum edge size of the image:

if ~isempty(opts.max_edge)

sz = size(im_rgb);

[~,max_dim] = max(sz(1:2));

osz = NaN*ones(1,2);

osz(max_dim) = opts.max_edge;

im_rgb = imresize(im_rgb, osz);

end

% do super-pixels:

fprintf(' > super-pix\n');

supix = gen_supperpixel_info(im_rgb, opts.sp_size);

pinfo = gen_feature_info_pairwise(im_rgb, supix);

% build "data-set":

ds=[];

ds.img_idxes = 1;

ds.img_data = im_rgb;

ds.sp_info{1} = supix;

ds.pws_info = pinfo;

ds.sp_num_imgs = supix.sp_num;

% run cnn:

fprintf(' > CNN\n');

depth = do_model_evaluate(model, ds, opts);

if opts.inpaint

fprintf(' > inpaint\n');

depth = do_inpainting(depth, im_rgb, supix);

end

if opts.normalize_depth

d_min = min(depth(:));

d_max = max(depth(:));

depth = (depth-d_min) / (d_max-d_min);

depth(depth<0) = 0;

depth(depth>1) = 1;

end