

In [1]:

    
import numpy as np
import os,sys
from PIL import Image
import re
import matplotlib.image as mpimg

from time import time
from  sklearn.ensemble import GradientBoostingRegressor
import mask_to_submission

Please indicate the path to the data



In [2]:

    
path_to_data = ""



In [3]:

    
# Loaded a set of images
root_dir = path_to_data + "training/"

image_dir = root_dir + "images/"
files = os.listdir(image_dir)[1:]
n = len(files)
print("Loading " + str(n) + " images")
imgs = np.array([mpimg.imread(image_dir + files[i]) for i in range(n)])

gt_dir = root_dir + "groundtruth/"
print("Loading " + str(n) + " images")
gt_imgs = np.array([mpimg.imread(gt_dir + files[i]) for i in range(n)])









    



Loading 100 images
Loading 100 images

Gradient boosting regression



In [4]:

    
x_train = np.array(imgs)
y_train = np.array(gt_imgs)


st = time()

model = GradientBoostingRegressor(n_estimators=500)
model.fit(x_train.reshape(-1,3),y_train.reshape(-1))

print time() - st









    



17.124311924

Testset's importation



In [5]:

    
root_dir = path_to_data + "test_set_images/"

n = 51
print("Loading " + str(n) + " images")
imgs = np.array([mpimg.imread(root_dir + "test_"+str(i)+"/" + "test_"+str(i)+".png") for i in range(1,n)])









    



Loading 51 images

classification prediction



In [6]:

    
pred_test = model.predict(imgs.reshape(-1, 3))
pred_test = pred_test.reshape(imgs.shape[:-1])

Segmentation with radom_walker



In [7]:

    
from skimage.segmentation import random_walker
from sklearn.metrics import f1_score
def pred_random_walker(reg, thr_non_road, thr_road, beta):
    '''
    Process random walker algorithm for every images.
    
    reg: np.ndarray(number of image, number of row,  number of column)
    thr_road: minimum value for being consider as a road
    thr_non_road: minimum value for being consider as a non-road
    beta: random walker's penalization
    '''
    segmentation = np.zeros_like(reg)
    for i in range(len(reg)):
        data = reg[i]
        
        markers=np.zeros_like(data)
        markers[data<thr_non_road] = 1
        markers[data>thr_road] = 2
        segmentation[i] = random_walker(data, markers, beta=beta, mode='bf')
    
    segmentation = segmentation -1
    return segmentation



In [14]:

    
image_shape = imgs.shape[:-1]
segmentations = pred_random_walker(pred_test.reshape(image_shape), 0.05, 0.5, beta=500)

Write the segmentations from the random walker algorithm in files.



In [15]:

    
# creation of the test masks folder if not exist
try:
    os.stat(path_to_data + 'test_set_masks')
except:
    os.mkdir(path_to_data + 'test_set_masks')



In [16]:

    
import scipy.misc
def write_segmentation( segmentations):
    for i in range(1, len(segmentations)+1):
        filename =  path_to_data + 'test_set_masks/satImage_' + '%.3d' % i + '.png'
        scipy.misc.imsave(filename, segmentations[i-1])



In [17]:

    
write_segmentation( segmentations)

make the submission file



In [18]:

    
submission_filename = 'baseline_submission.csv'
image_filenames = []

for i in range(1, 51):
    image_filename = path_to_data + 'test_set_masks/satImage_' + '%.3d' % i + '.png'
    image_filenames.append(image_filename)
mask_to_submission.masks_to_submission(submission_filename, *image_filenames)