def extract_signals_im_training(imageNameFile, maskNameFile, gtNameFile):
"""
Given the paths of Image (original), Mask (of the signals), and the GT (with the annotations)
it builds a signal list of the signals in the image
"""
image = cv.imread(imageNameFile)
imageMask = cv.imread(maskNameFile)
annotations = load_annotations(gtNameFile)
signal_list = []
for annotation in annotations:
signal = Signal()
signal.build_from_annotation(annotation, image, imageMask)
signal.img_orig_path = imageNameFile
signal.img_mask_path = maskNameFile
signal_list.append(signal)
return signal_list
pixelFP = pixelFP + localPixelFP
pixelFN = pixelFN + localPixelFN
pixelTN = pixelTN + localPixelTN
if window_evaluation == 1:
# Read .pkl file
name_r, ext_r = os.path.splitext(result_files[ii])
pkl_name = '{}/{}/{}.pkl'.format(results_dir, method, name_r)
with open(pkl_name, "rb") as fp: # Unpickling
windowCandidates = pickle.load(fp)
gt_annotations_name = '{}/gt/gt.{}.txt'.format(test_dir, name)
windowAnnotations = load_annotations(gt_annotations_name)
[localWindowTP, localWindowFN, localWindowFP] = evalf.performance_accumulation_window(windowCandidates, windowAnnotations)
windowTP = windowTP + localWindowTP
windowFN = windowFN + localWindowFN
windowFP = windowFP + localWindowFP
# Plot performance evaluation
[pixelPrecision, pixelAccuracy, pixelSpecificity, pixelSensitivity] = evalf.performance_evaluation_pixel(pixelTP, pixelFP, pixelFN, pixelTN)
pixelF1 = 0
if (pixelPrecision + pixelSensitivity) != 0:
pixelF1 = 2*((pixelPrecision*pixelSensitivity)/(pixelPrecision + pixelSensitivity))
print ('Team {:02d} pixel, method {} : {:.2f}, {:.2f}, {:.2f}\n'.format(team, method, pixelPrecision, pixelSensitivity, pixelF1))
if window_evaluation == 1:
def traffic_sign_detection(directory, output_dir, pixel_method, window_method):
pixelTP = 0
pixelFN = 0
pixelFP = 0
pixelTN = 0
windowTP = 0
windowFN = 0
windowFP = 0
window_precision = 0
window_accuracy = 0
# Load image names in the given directory
file_names = sorted(fnmatch.filter(os.listdir(directory), '*.jpg'))
pixel_time = 0
for name in file_names:
base, extension = os.path.splitext(name)
# Read file
im = imageio.imread('{}/{}'.format(directory, name))
print('{}/{}'.format(directory, name))
# Candidate Generation (pixel) ######################################
start = time.time()
pixel_candidates = candidate_generation_pixel(im, pixel_method)
end = time.time()
pixel_time += (end - start)
fd = '{}/{}_{}'.format(output_dir, pixel_method, window_method)
if not os.path.exists(fd):
os.makedirs(fd)
out_mask_name = '{}/{}.png'.format(fd, base)
imageio.imwrite(out_mask_name, np.uint8(np.round(pixel_candidates)))
if window_method != 'None':
window_candidates = candidate_generation_window(
im, pixel_candidates, window_method)
out_list_name = '{}/{}.pkl'.format(fd, base)
with open(out_list_name, "wb") as fp: #Pickling
pickle.dump(window_candidates, fp)
# Accumulate pixel performance of the current image #################
pixel_annotation = imageio.imread('{}/mask/mask.{}.png'.format(
directory, base)) > 0
[localPixelTP, localPixelFP, localPixelFN, localPixelTN
] = evalf.performance_accumulation_pixel(pixel_candidates,
pixel_annotation)
pixelTP = pixelTP + localPixelTP
pixelFP = pixelFP + localPixelFP
pixelFN = pixelFN + localPixelFN
pixelTN = pixelTN + localPixelTN
[
pixel_precision, pixel_accuracy, pixel_recall, pixel_specificity,
pixel_sensitivity, pixel_F1, pixel_TP, pixel_FP, pixel_FN
] = evalf.performance_evaluation_pixel(pixelTP, pixelFP, pixelFN,
pixelTN)
if window_method != 'None':
# Accumulate object performance of the current image ################
window_annotationss = load_annotations('{}/gt/gt.{}.txt'.format(
directory, base))
[localWindowTP, localWindowFN, localWindowFP
] = evalf.performance_accumulation_window(window_candidates,
window_annotationss)
windowTP = windowTP + localWindowTP
windowFN = windowFN + localWindowFN
windowFP = windowFP + localWindowFP
# Plot performance evaluation
[window_precision, window_sensitivity, window_accuracy
] = evalf.performance_evaluation_window(windowTP, windowFN,
windowFP)
pixel_time /= len(file_names)
return [
pixel_precision, pixel_accuracy, pixel_recall, pixel_specificity,
pixel_sensitivity, pixel_F1, pixel_TP, pixel_FP, pixel_FN, pixel_time
]
def visualizeHistograms(imPath, gtPath, maskPath, colorSpace="RGB"):
"""
Plots histograms in the color space selected of all the signals pixels values.
Main use for profiling mask margins
* Inputs:
- imPath = path to train images
- gtPath = path to annotations
- maskPath = path to masks
*Outputs:
- None
"""
from main import CONSOLE_ARGUMENTS
file_names = sorted(fnmatch.filter(os.listdir(imPath), '*.jpg'))
histAll = [[[0] for _ in range(3)] for _ in range(6)]
for name in file_names[:-1]:
base, extension = os.path.splitext(name)
imageNameFile = imPath + "/" + name
maskNameFile = maskPath + "/mask." + base + ".png"
gtNameFile = gtPath + "/gt." + base + ".txt"
image = cv.imread(imageNameFile)
if (CONSOLE_ARGUMENTS.histogram_norm):
image = cv.cvtColor(image, cv.COLOR_BGR2RGB)
image = preprocess_normrgb(image)
image = cv.cvtColor(image, cv.COLOR_RGB2BGR)
if (colorSpace == "LAB"):
image = cv.cvtColor(image, cv.COLOR_BGR2Lab)
if (colorSpace == "Luv"):
image = cv.cvtColor(image, cv.COLOR_BGR2Luv)
if (colorSpace == "normRGB"):
image = cv.cvtColor(image, cv.COLOR_BGR2RGB)
image = preprocess_normrgb(image)
elif (colorSpace == "HSL"):
image = cv.cvtColor(image, cv.COLOR_BGR2HLS)
elif (colorSpace == "HSV"):
image = cv.cvtColor(image, cv.COLOR_BGR2HSV)
elif (colorSpace == "Yuv"):
image = cv.cvtColor(image, cv.COLOR_BGR2YUV)
elif (colorSpace == "XYZ"):
image = cv.cvtColor(image, cv.COLOR_BGR2XYZ)
elif (colorSpace == "YCrCb"):
image = cv.cvtColor(image, cv.COLOR_BGR2YCrCb)
maskImage = cv.imread(maskNameFile)
image = image * maskImage
annot = load_annotations(gtNameFile)
for rect in annot:
tly, tlx, bly, blx = rect[:4]
tly, tlx, bly, blx = int(tly), int(tlx), int(bly), int(blx)
color = ('b', 'g', 'r')
for i in range(3):
histr = cv.calcHist(image[tly:bly, tlx:blx, :], [i], None,
[60], [0, 256])
histr[0] = 0
histAll[signal_dicts[rect[4]]][i] += histr
titles = ["A", "B", "C", "D", "E", "F"]
dir = CONSOLE_ARGUMENTS.hist_save_directories
for j, hist_signal_type in enumerate(histAll):
color = ('b', 'g', 'r')
plt.figure()
plt.title(titles[j])
plt.ioff()
for i, col in enumerate(color):
hist_signal_type[i][0] = 0
plt.plot(hist_signal_type[i], color=col)
plt.xlim([0, 60])
directory = dir + "/" + colorSpace
if not os.path.exists(directory):
os.makedirs(directory)
plt.savefig(directory + "/norm_" + titles[j] + ".png")
def traffic_sign_detection_test(directory, output_dir, pixel_method,
window_method):
"""
Calculates all statistical evaluation metrics of different pixel selector method (TRAINING AND VALIDATION)
* Inputs:
- directory = path to train images
- outpit_dir = Directory where to store output masks, etc. For instance '~/m1-results/week1/test'
- pixel_method = pixel method that will segmentate the image
- window_method = -------
*Outputs:
- pixel_precision, pixel_accuracy, pixel_specificity, pixel_sensitivity, window_precision, window_accuracy
"""
from main import CONSOLE_ARGUMENTS
pixelTP = 0
pixelFN = 0
pixelFP = 0
pixelTN = 0
windowTP = 0
windowFN = 0
windowFP = 0
window_precision = 0
window_accuracy = 0
print("splitting in trainning test")
# Load image names in the given directory
file_names = sorted(fnmatch.filter(os.listdir(directory), '*.jpg'))
signals_type_dict = get_dictionary()
training, validation = [], []
for key in signals_type_dict:
sig_subdict = signals_type_dict[key]
training_type, validation_type = divide_training_validation_SL(
sig_subdict['signal_list'])
training.extend(training_type)
validation.extend(validation_type)
print("extracting mask")
dataset = training
if (CONSOLE_ARGUMENTS.use_validation):
dataset = validation
# if(CONSOLE_ARGUMENTS.use_test):
totalTime = 0
for signal in dataset:
signal_path = signal.img_orig_path
_, name = signal_path.rsplit('/', 1)
base, extension = os.path.splitext(name)
# Read file
im = imageio.imread('{}/{}'.format(directory, name))
print('{}/{}'.format(directory, name))
# Candidate Generation (pixel) ######################################
start = time.time()
pixel_candidates = candidate_generation_pixel(im, pixel_method)
totalTime += time.time() - start
fd = '{}/{}_{}'.format(output_dir, pixel_method, window_method)
if not os.path.exists(fd):
os.makedirs(fd)
out_mask_name = '{}/{}.png'.format(fd, base)
imageio.imwrite(out_mask_name, np.uint8(np.round(pixel_candidates)))
if window_method != 'None':
window_candidates = candidate_generation_window(
im, pixel_candidates, window_method)
# Accumulate pixel performance of the current image #################
pixel_annotation = imageio.imread('{}/mask/mask.{}.png'.format(
directory, base)) > 0
[localPixelTP, localPixelFP, localPixelFN, localPixelTN
] = evalf.performance_accumulation_pixel(pixel_candidates,
pixel_annotation)
pixelTP = pixelTP + localPixelTP
pixelFP = pixelFP + localPixelFP
pixelFN = pixelFN + localPixelFN
pixelTN = pixelTN + localPixelTN
[
pixel_precision, pixel_accuracy, pixel_specificity,
pixel_sensitivity
] = evalf.performance_evaluation_pixel(pixelTP, pixelFP, pixelFN,
pixelTN)
if window_method != 'None':
# Accumulate object performance of the current image ################
window_annotationss = load_annotations('{}/gt/gt.{}.txt'.format(
directory, base))
[localWindowTP, localWindowFN, localWindowFP
] = evalf.performance_accumulation_window(window_candidates,
window_annotationss)
windowTP = windowTP + localWindowTP
windowFN = windowFN + localWindowFN
windowFP = windowFP + localWindowFP
# Plot performance evaluation
[window_precision, window_sensitivity, window_accuracy
] = evalf.performance_evaluation_window(windowTP, windowFN,
windowFP)
print("meanTime", totalTime / len(dataset))
print("pixelTP", pixelTP, "\t", pixelFP, "\t", pixelFN)
return [
pixel_precision, pixel_accuracy, pixel_specificity, pixel_sensitivity,
window_precision, window_accuracy
]
def traffic_sign_detection(directory, output_dir, pixel_method, window_method,
calculate_metrics):
pixelTP = 0
pixelFN = 0
pixelFP = 0
pixelTN = 0
pixel_F1 = 0
windowTP = 0
windowFN = 0
windowFP = 0
window_precision = 0
window_accuracy = 0
window_F1 = 0
counter = 0
# Load image names in the given directory
file_names = sorted(fnmatch.filter(os.listdir(directory), '*.jpg'))
for name in file_names:
counter += 1
base, extension = os.path.splitext(name)
# Read file
im = imageio.imread('{}/{}'.format(directory, name))
#print ('{}/{}'.format(directory,name))
# Candidate Generation (pixel) ######################################
pixel_candidates = candidate_generation_pixel(im, pixel_method)
# pixel_candidates = morphological_operators(pixel_candidates)
pixel_candidates = connected_labels_pixel_cand(im, pixel_candidates)
fd = '{}/{}_{}'.format(output_dir, pixel_method, window_method)
if not os.path.exists(fd):
os.makedirs(fd)
out_mask_name = '{}/{}.png'.format(fd, base)
if window_method != 'None':
window_candidates = candidate_generation_window(
im, pixel_candidates, window_method)
window_mask = np.zeros(pixel_candidates.shape)
for window_candidate in window_candidates:
window_mask[window_candidate[0]:window_candidate[2],
window_candidate[1]:
window_candidate[3]] = pixel_candidates[
window_candidate[0]:window_candidate[2],
window_candidate[1]:window_candidate[3]]
out_list_name = '{}/{}.pkl'.format(fd, base)
pixel_candidates = window_mask
with open(out_list_name, "wb") as fp: #Pickling
pickle.dump(window_candidates, fp)
imageio.imwrite(out_mask_name, np.uint8(np.round(pixel_candidates)))
pixel_precision = 0
pixel_accuracy = 0
pixel_specificity = 0
pixel_sensitivity = 0
window_precision = 0
window_accuracy = 0
if (calculate_metrics):
# Accumulate pixel performance of the current image #################
pixel_annotation = imageio.imread('{}/mask/mask.{}.png'.format(
directory, base)) > 0
[localPixelTP, localPixelFP, localPixelFN, localPixelTN
] = evalf.performance_accumulation_pixel(pixel_candidates,
pixel_annotation)
pixelTP = pixelTP + localPixelTP
pixelFP = pixelFP + localPixelFP
pixelFN = pixelFN + localPixelFN
pixelTN = pixelTN + localPixelTN
# [pixel_precision, pixel_accuracy, pixel_specificity, pixel_sensitivity, pixel_F1] = evalf.performance_evaluation_pixel(pixelTP, pixelFP, pixelFN, pixelTN)
if window_method != 'None':
# Accumulate object performance of the current image ################
window_annotationss = load_annotations(
'{}/gt/gt.{}.txt'.format(directory, base))
[localWindowTP, localWindowFN,
localWindowFP] = evalf.performance_accumulation_window(
window_candidates, window_annotationss)
windowTP = windowTP + localWindowTP
windowFN = windowFN + localWindowFN
windowFP = windowFP + localWindowFP
# Plot performance evaluation
# [window_precision, window_sensitivity, window_accuracy, window_F1] = evalf.performance_evaluation_window(windowTP, windowFN, windowFP)
if (calculate_metrics):
[
pixel_precision, pixel_accuracy, pixel_specificity,
pixel_sensitivity, pixel_F1
] = evalf.performance_evaluation_pixel(pixelTP, pixelFP, pixelFN,
pixelTN)
print("Pixel precision: " + str(pixel_precision))
print("Pixel accuracy: " + str(pixel_accuracy))
print("Pixel recall: " + str(pixel_sensitivity))
print("Pixel F1-measure: " + str(pixel_F1))
print("Pixel TP: " + str(pixelTP))
print("Pixel FP: " + str(pixelFP))
print("Pixel FN: " + str(pixelFN))
print("Pixel TN: " + str(pixelTN))
if window_method != 'None':
[window_precision, window_sensitivity, window_accuracy, window_F1
] = evalf.performance_evaluation_window(windowTP, windowFN,
windowFP)
print("Window precision: " + str(window_precision))
print("Window accuracy: " + str(window_accuracy))
print("Window F1-measure: " + str(window_F1))
return [
pixel_precision, pixel_accuracy, pixel_specificity, pixel_sensitivity,
pixel_F1, window_precision, window_accuracy, window_F1, counter
]
def traffic_sign_detection_test(directory, output_dir, pixel_method, window_method, use_dataset="training"):
"""
Calculates all statistical evaluation metrics of different pixel selector method (TRAINING AND VALIDATION)
* Inputs:
- directory = path to train images
- outpit_dir = Directory where to store output masks, etc. For instance '~/m1-results/week1/test'
- pixel_method = pixel method that will segmentate the image
- window_method = -------
*Outputs:
- pixel_precision, pixel_accuracy, pixel_specificity, pixel_sensitivity, window_precision, window_accuracy
"""
pixelTP = 0
pixelFN = 0
pixelFP = 0
pixelTN = 0
windowTP = 0
windowFN = 0
windowFP = 0
window_precision = 0
window_accuracy = 0
window_sensitivity = 0
# print("splitting in trainning test")
# Load image names in the given directory
# file_names = sorted(fnmatch.filter(os.listdir(directory), '*.jpg'))
signals_type_dict = get_dictionary()
training, validation = [], []
for key in signals_type_dict:
sig_subdict = signals_type_dict[key]
training_type, validation_type = divide_training_validation_SL(sig_subdict['signal_list'])
training.extend(training_type)
validation.extend(validation_type)
# print("extracting mask")
dataset = training
if(use_dataset == 'validation'):
dataset = validation
# if(CONSOLE_ARGUMENTS.use_test):
totalTime = 0
dataset_paths = [signal.img_orig_path for signal in dataset]
for signal_path in tqdm(dataset_paths, ascii=True, desc="Calculating Statistics"):
startTime = time.time()
rgb_mask, bb_list = get_pixel_candidates(signal_path)
totalTime = time.time() - startTime
if(bb_list is not None): bb_list = convertBBFormat(bb_list)
_, name = signal_path.rsplit('/', 1)
base, extension = os.path.splitext(name)
# Accumulate pixel performance of the current image #################
pixel_annotation = imageio.imread('{}/mask/mask.{}.png'.format(directory,base)) > 0
[localPixelTP, localPixelFP, localPixelFN, localPixelTN] =\
evalf.performance_accumulation_pixel(rgb_mask, pixel_annotation)
pixelTP = pixelTP + localPixelTP
pixelFP = pixelFP + localPixelFP
pixelFN = pixelFN + localPixelFN
pixelTN = pixelTN + localPixelTN
[pixel_precision, pixel_accuracy, pixel_specificity, pixel_sensitivity] =\
evalf.performance_evaluation_pixel(pixelTP, pixelFP, pixelFN, pixelTN)
if(bb_list != None):
# Accumulate object performance of the current image ################
window_annotationss = load_annotations('{}/gt/gt.{}.txt'.format(directory, base))
[localWindowTP, localWindowFN, localWindowFP] = \
evalf.performance_accumulation_window(bb_list, window_annotationss)
windowTP = windowTP + localWindowTP
windowFN = windowFN + localWindowFN
windowFP = windowFP + localWindowFP
# Plot performance evaluation
[window_precision, window_sensitivity, window_accuracy] = \
evalf.performance_evaluation_window(windowTP, windowFN, windowFP)
print("meanTime", totalTime/len(dataset))
print("pixelTP", pixelTP, "\t", pixelFP, "\t", pixelFN)
print("windowTP", windowTP, "\t", windowFP, "\t", windowFN)
return [pixel_precision, pixel_accuracy, pixel_specificity, pixel_sensitivity,\
window_precision, window_accuracy, window_sensitivity]