def __getitem__(self, index):
img_path = self.img_files[index]
# label_path = self.label_files[index]
model_stride = 2**4
labfile = splitext(img_path)[0] + '.txt'
L = readShapes(labfile)
img = cv2.imread(img_path) # BGR
assert img is not None, 'File Not Found ' + img_path
data = [img, L[0]]
img, labels = process_data_item(data, self.img_size, model_stride)
# labels = []
# if os.path.isfile(label_path):
# with open(label_path, 'r') as file:
# for line in file:
# data = line.strip().split(',')
# labels = data[1:9]
# text = data[9]
# Normalize
img = img[:, :, ::-1] - np.zeros_like(img, dtype=np.float32)
img = img.transpose(2, 0, 1)
# BGR to RGB, to 3x416x416
# img = np.ascontiguousarray(img, dtype=np.float32) # uint8 to float32
# img /= 255.0 # 0 - 255 to 0.0 - 1.0
# labels = np.array(labels).reshape(2,4)
labels = np.ascontiguousarray(labels, dtype=np.float32)
labels = torch.from_numpy(labels)
# labels = labels.requires_grad_()
img = torch.from_numpy(img)
# img = img.requires_grad_()
return img, labels #, (h, w)
def genOutput(img_path, output_dir, bname):
I = cv2.imread(img_path)
detected_cars_labels = '%s/%s_cars.txt' % (output_dir, bname)
Lcar = lread(detected_cars_labels)
sys.stdout.write('%s' % bname)
if Lcar:
for i, lcar in enumerate(Lcar):
draw_label(I, lcar, color=YELLOW, thickness=3)
lp_label = '%s/%s_%dcar_lp.txt' % (output_dir, bname, i)
lp_label_str = '%s/%s_%dcar_lp_str.txt' % (output_dir, bname, i)
if isfile(lp_label):
Llp_shapes = readShapes(lp_label)
pts = Llp_shapes[0].pts * lcar.wh().reshape(
2, 1) + lcar.tl().reshape(2, 1)
ptspx = pts * np.array(I.shape[1::-1], dtype=float).reshape(
2, 1)
draw_losangle(I, ptspx, RED, 3)
if isfile(lp_label_str):
with open(lp_label_str, 'r') as f:
lp_str = f.read().strip()
llp = Label(0, tl=pts.min(1), br=pts.max(1))
write2img(I, llp, lp_str)
sys.stdout.write(',%s' % lp_str)
cv2.imwrite('%s/%s_output.png' % (output_dir, bname), I,
[int(cv2.IMWRITE_PNG_COMPRESSION), 9])
def sample_validate(image_abs_path, gt_file):
L = readShapes(gt_file)[0]
pts_transformed = L.pts
model_stride = 16.0
image_ndarray = cv2.imread(image_abs_path)
tl, br = pts_transformed.min(1), pts_transformed.max(1)
llp = Label(0, tl, br)
YY, is_there_plate = labels2output_map(llp, pts_transformed, 304, 304,
model_stride, L.cls, image_abs_path)
print('There is place: %s | Class %s ' % (str(is_there_plate), L.cls))
print('Car ')
print(np.matrix(YY[..., 0]))
print('Non-object')
print(np.matrix(YY[..., 10]))
print('Moto')
print(np.matrix(YY[..., 9]))
def generate_outputs(input_dir, output_dir):
img_files = image_files_from_folder(input_dir)
for img_file in img_files:
bname = splitext(basename(img_file))[0]
I = cv2.imread(img_file)
detected_cars_labels = '%s/%s_cars.txt' % (output_dir, bname)
Lcar = lread(detected_cars_labels)
sys.stdout.write('%s' % bname)
if Lcar:
for i, lcar in enumerate(Lcar):
draw_label(I, lcar, color=YELLOW, thickness=3)
lp_label = '%s/%s_%dcar_lp.txt' % (output_dir, bname, i)
lp_label_str = '%s/%s_%dcar_lp_str.txt' % (output_dir, bname, i)
if isfile(lp_label):
Llp_shapes = readShapes(lp_label)
pts = Llp_shapes[0].pts * lcar.wh().reshape(2, 1) + lcar.tl().reshape(2, 1)
ptspx = pts * np.array(I.shape[1::-1], dtype=float).reshape(2, 1)
draw_losangle(I, ptspx, RED, 3)
if isfile(lp_label_str):
with open(lp_label_str, 'r') as f:
lp_str = f.read().strip()
llp = Label(0, tl=pts.min(1), br=pts.max(1))
write2img(I, llp, lp_str)
sys.stdout.write(',%s' % lp_str)
cv2.imwrite('%s/%s_output.png' % (output_dir, bname), I)
def labelImgShow(): YELLOW = ( 0,255,255) RED = ( 0, 0,255) img_file = '/home/yaokun/code/alpr-unconstrained/samples/train-detector/00011.jpg' lp_label = '/home/yaokun/code/alpr-unconstrained/samples/train-detector/00011.txt' I = cv2.imread(img_file) if isfile(lp_label): #pt1 = tuple(pts[:,i].astype(int).tolist()) #pt2 = tuple(pts[:,(i+1)%4].astype(int).tolist()) #cv2.line(I,pt1,pt2,color,thickness) Llp_shapes = readShapes(lp_label) print(np.array(I.shape[1::-1]).reshape(2,1)) print(Llp_shapes[0].pts) pts = Llp_shapes[0].pts*np.array(I.shape[1::-1],dtype=float).reshape(2,1) print(pts) #ptspx = pts*np.array(I.shape[1::-1],dtype=float).reshape(2,1) draw_losangle(I,pts,RED,3) show(I)
key_roll_decrease = LESS_THAN
key_perspective_reset = HOME
zoom_factor = 1.
action_keys = [key_exit, key_previous] + key_next
curr_image = 0
while curr_image < len(img_files):
img_file = img_files[curr_image]
lab_file = splitext(img_file)[0] + '.txt'
if isfile(lab_file):
shapes = readShapes(lab_file, obj_type=ShapeDisplay)
selected = len(shapes) - 1
else:
shapes, selected = [ShapeDisplay()], 0
zoom_factor = 1.
disp = Display(cv2.imread(img_file), maxW, maxH)
disp.show()
key = disp.waitKey()
typing_mode = False
while not key in action_keys:
disp.resetDisplay()
displayAllShapes(disp, shapes, selected, typing_mode)
disp.show()
if not isdir(outdir):
makedirs(outdir)
model, model_stride, xshape, yshape = load_network(args.model, dim)
opt = getattr(keras.optimizers, args.optimizer)(lr=args.learning_rate)
model.compile(loss=loss, optimizer=opt)
print('Checking input directory...')
Files = image_files_from_folder(train_dir)
Data = []
for file in Files:
labfile = splitext(file)[0] + '.txt'
if isfile(labfile):
L = readShapes(labfile)
I = cv2.imread(file)
Data.append([I, L[0]])
print('%d images with labels found' % len(Data))
dg = DataGenerator( data=Data, \
process_data_item_func=lambda x: process_data_item(x,dim,model_stride),\
xshape=xshape, \
yshape=(yshape[0],yshape[1],yshape[2]+1), \
nthreads=2, \
pool_size=1000, \
min_nsamples=100 )
dg.start()
Xtrain = np.empty((batch_size, dim, dim, 3), dtype='single')
Lcar = lread(detected_cars_labels)
sys.stdout.write('%s' % bname)
if Lcar:
for i, lcar in enumerate(Lcar):
draw_label(I, lcar, color=YELLOW, thickness=3)
lp_label = '%s/%s_%dcar_lp.txt' % (output_dir, bname, i)
lp_label_str = '%s/%s_%dcar_lp_str.txt' % (output_dir, bname, i)
if isfile(lp_label):
Llp_shapes = readShapes(lp_label)
pts = Llp_shapes[0].pts * lcar.wh().reshape(
2, 1) + lcar.tl().reshape(2, 1)
ptspx = pts * np.array(I.shape[1::-1], dtype=float).reshape(
2, 1)
draw_losangle(I, ptspx, RED, 3)
if isfile(lp_label_str):
with open(lp_label_str, 'r', encoding='utf-8') as f:
lp_str = f.read().strip()
llp = Label(0, tl=pts.min(1), br=pts.max(1))
I = write2img(I, llp, lp_str)
sys.stdout.write(',%s' % lp_str)
cv2.imwrite('%s/%s_output.png' % (output_dir, bname), I)
file_gr_transformed.close()
salvar_imagem(img_ndarray, caminho_completo_saida)
# if not possui_bbox:
# nome_arquivo = 'BAD_SAMPLE_'+nome_arquivo
# caminho_completo_saida_bbox = os.path.abspath(os.path.join(diretorio_saida, nome_arquivo))
def salvar_imagem(imagem_array, caminho_completo_arquivo):
cv2.imwrite(caminho_completo_arquivo, imagem_array)
if __name__ == '__teste3__':
data_item = []
# L = readShapes('/media/jones/dataset/alpr/lotes_rotulacao/preprocessados/train/lote1_1703_2000000544.txt')
# file = '/media/jones/dataset/alpr/lotes_rotulacao/preprocessados/train/lote1_1703_2000000544.jpg'
L = readShapes('samples/034_1304_1354_1705_0000_0130m000000041.txt')
file = 'samples/034_1304_1354_1705_0000_0130m000000041.jpg'
# I = cv2.imread(file)
dim_w = 304
dim_h = 304
# data_item.append([file, L[0]])
data_item.append(file)
data_item.append(L[0])
model_stride = 16.0
nome_arquivo = '034_1304_1354_1705_0000_0130m000000041'
nome_arquivo_imagem = 'samples/' + nome_arquivo + '_transf.jpg'
nome_arquivo_gt = 'samples/' + nome_arquivo + '_transf.txt'
XX, YY_output_map, pts_transformed = process_data_item(
data_item, dim_w, dim_h, model_stride)
sample_height, sample_width = XX.shape[:2]
# x_points = (pts_transformed[0:]*sample_width).tolist()
batch_size = args.batch_size
dim = 304
model,model_stride,xshape,yshape = load_network(args.model,dim)
opt = getattr(keras.optimizers,args.optimizer)(lr=args.learning_rate)
model.compile(loss=loss_mc, optimizer=opt)
# model.compile(loss=loss, optimizer=opt)
print('yshape: %s ' % str(yshape))
for layer in model.layers:
print(layer.name)
print(layer.output_shape)
print('Checking input directory...')
Files = image_files_from_folder(train_dir)
dataset_samples = []
for file in Files:
# file = Files[0]
labfile = splitext(file)[0] + '.txt'
gt_list = readShapes(labfile)
for gt_sample in gt_list:
dataset_samples.append([file, gt_sample])
XX, YY, pts_transformed = process_data_item(dataset_samples[0], dim, dim, model_stride)
Xtrain = np.empty((batch_size, dim, dim, 3), dtype='single')
Ytrain = np.empty((batch_size, int(dim / model_stride), int(dim / model_stride), 2 * 4 + 3))
Xtrain[0] = XX
Ytrain[0] = YY
XX, YY, pts_transformed = process_data_item(dataset_samples[1], dim, dim, model_stride)
Xtrain[1] = XX
Ytrain[1] = YY
print('starting train')
train_loss = model.train_on_batch(Xtrain, Ytrain)
print('loss %f ' % train_loss)
# print(flags)
