def main():
##
ext = '.png'
##
path, txt_file, path_converted = process_arguments(sys.argv)
# Create dir for converted labels
if not os.path.isdir(path_converted):
os.makedirs(path_converted)
with open(txt_file, 'rb') as f:
for img_name in f:
img_base_name = img_name.strip()
img_base_name = str(img_base_name, encoding='utf-8')
print(path)
print(img_base_name)
img_name = os.path.join(path, img_base_name) + ext
img = imread(img_name)
if (len(img.shape) > 2):
img = convert_from_color_segmentation(img)
imsave(os.path.join(path_converted, img_base_name) + ext, img)
else:
print(img_name +
" is not composed of three dimensions, therefore "
"shouldn't be processed by this script.\n"
"Exiting.",
file=sys.stderr)
exit()
def outgenf(img_name):
Prediction_dir = "../../data/Prediction"
Prediction_name = str(img_name) +".png"
Prediction_path = os.path.join(Prediction_dir,Prediction_name)
prediction_img = imread(Prediction_path)
prediction = utils.convert_from_color_segmentation(prediction_img)
return prediction
def preprocess_label(img, lut, mode, im_sz):
# If label is three-dimensional image we have to convert it to
# corresponding labels (0 - 20). Currently anticipated labels are from
# VOC pascal datasets.
if (len(img.shape) > 2):
img = convert_from_color_segmentation(img)
img = preprocess_data(img, mode, im_sz, 'label')
img = lut[img]
img = np.expand_dims(img, axis=0)
#img = _2D_to_ND(img, len(np.unique(lut)))
#img = img.transpose((2,0,1))
return img
def contain_class(img_name, class_ids, class_names):
img = imread(img_name)
# If label is three-dimensional image we have to convert it to
# corresponding labels (0 - 20). Currently anticipated labels are from
# VOC pascal datasets.
if (len(img.shape) > 2):
img = convert_from_color_segmentation(img)
for i,j in enumerate(class_ids):
if j in np.unique(img):
return class_names[i]
return False
def contain_class(img_name, class_ids, lut):
img = imread(img_name)
# If label is three-dimensional image we have to convert it to
# corresponding labels (0 - 20). Currently anticipated labels are from
# VOC pascal datasets.
if (len(img.shape) > 2):
img = convert_from_color_segmentation(img)
img_labels = np.unique(img)
if len(set(img_labels).intersection(class_ids)) >= 1:
return lut[img]
else:
return None
def contain_class(img_name, class_ids, class_names):
'''
arguments:
img_name: name of image
class_ids: interested classes index
class_names: interested classes name
'''
img = imread(img_name)
# If label is three-dimensional image we have to convert it to
# corresponding labels (0 - 20). Currently anticipated labels are from
# VOC pascal datasets.
# if img is rgb structure, transform img grayscale
if (len(img.shape) > 2):
img = convert_from_color_segmentation(img)
for i, j in enumerate(class_ids):
if j in np.unique(img):
# if image pixel have class_id, return calss name.
return class_names[i]
return False
def evaluate_IoU_class_general(truth_dir, outgenf, imglist_file):
class_count = len(
utils.pascal_classes()) + 1 #include background as a class
int_vec = [list() for i in range(class_count)]
union_vec = [list() for i in range(class_count)]
with open(imglist_file, "r") as imf:
im_names = imf.read().splitlines()
im_counts = len(im_names)
#Store the IoU for each specific image
iou_img_dict = {}
int_vec = np.zeros((class_count, im_counts))
union_vec = np.zeros((class_count, im_counts))
ptr = 0
for img_name in im_names:
img_truth_name = os.path.join(truth_dir, img_name) + '.png'
img_truth = imread(img_truth_name)
truth_data = utils.convert_from_color_segmentation(img_truth)
output_data = outgenf(img_name)
#Calculate IoU for each class
iou_img = np.zeros(class_count)
for class_idx in range(0, class_count):
int_i, uni_i = int_uni_cls(truth_data, output_data, class_idx)
if (int_i == 0 and uni_i == 0):
iou_img[class_idx] = np.NaN
else:
iou_img[class_idx] = int_i / uni_i
int_vec[class_idx, ptr] = int_i
union_vec[class_idx, ptr] = uni_i
iou_img_dict[
"img_" +
img_name] = iou_img #add "img_" here to ensure variable name is valid in matlab
ptr = ptr + 1
iou_cls = [
np.sum(int_vec[i]) / np.sum(union_vec[i]) for i in range(class_count)
]
return iou_cls, iou_img_dict
with open(Name_list, "r") as f:
lines = f.read().splitlines()
np.random.seed(42)
lines_shuffled = np.random.permutation(lines)
Sample_size = 94
lines_sampled = lines_shuffled[0:Sample_size]
Max_iteration = 20
kl_divergence = np.zeros((Sample_size, Max_iteration))
for i, line in enumerate(lines_sampled):
Image_name = str(line) + ".jpg"
Image_path = os.path.join(Image_dir, Image_name)
Unary_name = str(line) + ".mat"
Unary_path = os.path.join(Unary_dir, Unary_name)
img_truth_name = os.path.join(Ground_truth_dir, str(line)) + '.png'
img_truth = imread(img_truth_name)
truth_data = utils.convert_from_color_segmentation(img_truth)
img = imread(Image_path)
unary = scipy.io.loadmat(Unary_path)
unary_energy = unary["energy"].reshape(21, img.shape[0] * img.shape[1])
crf = DenseCRF(img.shape[0], img.shape[1], 21, Ground_truth=truth_data)
crf.set_unary(-unary_energy)
crf.add_pairwise_gaussian(1, 1, PottsCompatibility(3))
crf.add_pairwise_bilateral(61, 61, 11, 11, 11, img, PottsCompatibility(10))
Q = crf.inference(Max_iteration)
kl_divergence[i, :] = crf.KL_divergence
print(kl_divergence[i, :])
scipy.io.savemat(os.path.join(output_path, "kl_divergence"),
{"kl_divergence": kl_divergence})
