dataset_setup.m

%%
%
% This is a script for reading the labelme annotations and outputing a
% dataset that can be used for training a CNN. It is specifically designed
% to work with KittiSeg network which uses annotation images with color
% masks denoting the labels.
%
%#ok<*AGROW>
%
%% Read the images and annotations

clear;close all;clc;


ANNOTATION_DIR = '../collection/Annotations/';
IMAGE_DIR = '../collection/Images/';
OUTPUT_DIR = '../dataset/ground_truth/';


[D, XML] = LMdatabase(ANNOTATION_DIR);

%% Filter the images not annotated


% filter all images without any obects (ie. ones that have not been
% annotated)
database = [];
for i = 1 : length(D)
    if isfield(D(i).annotation, 'object')
        database = [database; D(i).annotation];
    end
end

%% Parse annotations and save annotated images

% red and purple are the default colors

for j = 1 : length(database)
    [mask, class] = LMobjectmask(database(j), IMAGE_DIR);
    [r,c,~] = size(mask);
    mask_composite = false(r,c);
    for i = 1 : length(class)
        if strcmp(class(i),'line') || strcmp(class(i),'pothole')
            mask_composite = mask_composite | mask(:,:,i);
        end
    end
    annotation_image = cat(3, 255*ones(r,c), zeros(r,c,2));
    blue_mask = cat(3, false(r,c,2), mask_composite);
    annotation_image(blue_mask) = annotation_image(blue_mask) + 255;
    imwrite(annotation_image, strcat(OUTPUT_DIR, database(j).filename));
end