def predict(args):
# complie models
model = nn.network()
# model.load_weights('3dbox_weights_1st.hdf5')
model.load_weights(args.w)
# KITTI_train_gen = KITTILoader(subset='training')
dims_avg, _ = KITTILoader(subset='training').get_average_dimension()
# list all the validation images
if args.a == 'training':
all_imgs = sorted(os.listdir(test_image_dir))
val_index = int(len(all_imgs) * cfg().split)
val_imgs = all_imgs[val_index:]
else:
val_imgs = sorted(os.listdir(test_image_dir))
start_time = time.time()
for i in val_imgs:
image_file = test_image_dir + i
label_file = test_label_dir + i.replace('png', 'txt')
prediction_file = prediction_path + i.replace('png', 'txt')
calibration_file = test_calib_path + i.replace('png', 'txt')
# write the prediction file
with open(prediction_file, 'w') as predict:
img = cv2.imread(image_file)
img = np.array(img, dtype='float32')
P2 = np.array([])
for line in open(calibration_file):
if 'P2' in line:
P2 = line.split(' ')
P2 = np.asarray([float(i) for i in P2[1:]])
P2 = np.reshape(P2, (3, 4))
for line in open(label_file):
line = line.strip().split(' ')
obj = detectionInfo(line)
xmin = int(obj.xmin)
xmax = int(obj.xmax)
ymin = int(obj.ymin)
ymax = int(obj.ymax)
if obj.name in cfg().KITTI_cat:
# cropped 2d bounding box
if xmin == xmax or ymin == ymax:
continue
# 2D detection area
patch = img[ymin:ymax, xmin:xmax]
patch = cv2.resize(patch, (cfg().norm_h, cfg().norm_w))
patch -= np.array([[[103.939, 116.779, 123.68]]])
# extend it to match the training dimension
patch = np.expand_dims(patch, 0)
prediction = model.predict(patch)
dim = prediction[0][0]
bin_anchor = prediction[1][0]
bin_confidence = prediction[2][0]
# update with predict dimension
dims = dims_avg[obj.name] + dim
obj.h, obj.w, obj.l = np.array(
[round(dim, 2) for dim in dims])
# update with predicted alpha, [-pi, pi]
obj.alpha = recover_angle(bin_anchor, bin_confidence,
cfg().bin)
# compute global and local orientation
obj.rot_global, rot_local = compute_orientaion(P2, obj)
# compute and update translation, (x, y, z)
obj.tx, obj.ty, obj.tz = translation_constraints(
P2, obj, rot_local)
# output prediction label
output_line = obj.member_to_list()
output_line.append(1.0)
# Write regressed 3D dim and orientation to file
output_line = ' '.join([str(item)
for item in output_line]) + '\n'
predict.write(output_line)
print('Write predicted labels for: ' + str(i))
end_time = time.time()
process_time = (end_time - start_time) / len(val_imgs)
print(process_time)
def predict(prediction_dir, label_dir, image_dir, calibration_file):
# complie models
model = nn.network()
model.load_weights('3dbox_weights_mob.hdf5')
# model.load_weights(args.w)
# KITTI_train_gen = KITTILoader(subset='training')
dims_avg, _ = KITTILoader(subset='tracklet').get_average_dimension()
val_imgs = sorted(
[im for im in os.listdir(image_dir) if not im.startswith('.')])
P2 = np.array([])
for line in open(calibration_file):
if 'P2' in line:
P2 = line.split(' ')
P2 = np.asarray([float(i) for i in P2[1:]])
P2 = np.reshape(P2, (3, 4))
for img in tqdm(val_imgs):
image_file = os.path.join(image_dir, img)
label_file = os.path.join(label_dir, img.replace('png', 'txt'))
prediction_file = os.path.join(prediction_dir,
img.replace('png', 'txt'))
# write the prediction file
with open(prediction_file, 'w') as predict:
img = cv2.imread(image_file)
img = np.array(img, dtype='float32')
for line in open(label_file):
line = line.strip().split(' ')
obj = detectionInfo(line)
xmin = int(obj.xmin)
xmax = int(obj.xmax)
ymin = int(obj.ymin)
ymax = int(obj.ymax)
if obj.name in cfg.KITTI_cat:
# cropped 2d bounding box
if xmin == xmax or ymin == ymax:
continue
# 2D detection area
patch = img[ymin:ymax, xmin:xmax]
try:
patch = cv2.resize(patch, (cfg.norm_h, cfg.norm_w))
except cv2.error:
continue
# patch -= np.array([[[103.939, 116.779, 123.68]]])
patch /= 255.0
# extend it to match the training dimension
patch = np.expand_dims(patch, 0)
prediction = model.predict(patch)
dim = prediction[0][0]
bin_anchor = prediction[1][0]
bin_confidence = prediction[2][0]
# update with predict dimension
dims = dims_avg[obj.name] + dim
obj.h, obj.w, obj.l = np.array(
[round(dim, 2) for dim in dims])
# update with predicted alpha, [-pi, pi]
obj.alpha = recover_angle(bin_anchor, bin_confidence,
cfg.bin)
# compute global and local orientation
obj.rot_global, rot_local = compute_orientaion(P2, obj)
# compute and update translation, (x, y, z)
obj.tx, obj.ty, obj.tz = translation_constraints(
P2, obj, rot_local)
# output prediction label
output_line = obj.member_to_list()
output_line.append(1.0)
# Write regressed 3D dim and orientation to file
output_line = ' '.join([str(item)
for item in output_line]) + '\n'
predict.write(output_line)
def train():
KITTI_train_gen = KITTILoader(subset='training')
dim_avg, dim_cnt = KITTI_train_gen.get_average_dimension()
new_data = orientation_confidence_flip(KITTI_train_gen.image_data, dim_avg)
model = nn.network()
#model.load_weights('model00000296.hdf5')
early_stop = callbacks.EarlyStopping(monitor='val_loss', min_delta=0.001, patience=10, mode='min', verbose=1)
checkpoint = callbacks.ModelCheckpoint('model{epoch:08d}.hdf5', monitor='val_loss', verbose=1, save_best_only=False, mode='min', period=1)
tensorboard = callbacks.TensorBoard(log_dir='logs/', histogram_freq=0, write_graph=True, write_images=False)
all_examples = len(new_data)
trv_split = int(cfg().split * all_examples) # train val split
train_gen = data_gen(new_data[: trv_split])
valid_gen = data_gen(new_data[trv_split : all_examples])
print("READY FOR TRAINING")
train_num = int(np.ceil(trv_split / cfg().batch_size))
valid_num = int(np.ceil((all_examples - trv_split) / cfg().batch_size))
#gen_flow = gen_flow_for_two_inputs(X_train, X_angle_train, y_train)
# choose the minimizer to be sgd
# minimizer = optimizer.SGD(lr=0.0001, momentum = 0.9)
minimizer = optimizer.Adam(lr=0.0001)
# multi task learning
model.compile(optimizer=minimizer, #minimizer,
loss={'dimensions': 'mean_squared_error', 'orientation': orientation_loss, 'confidence': 'categorical_crossentropy'},
loss_weights={'dimensions': 1., 'orientation': 10., 'confidence': 5.})
print("####################################################")
print(K.get_value(model.optimizer.lr))
# Tambahan aing
def scheduler(epoch):
if epoch%10==0 and epoch!=0:
lr = K.get_value(model.optimizer.lr)
K.set_value(model.optimizer.lr, lr*.8)
print("lr changed to {}".format(lr*.8))
print("lr = ", K.get_value(model.optimizer.lr))
return K.get_value(model.optimizer.lr)
lr_sched = callbacks.LearningRateScheduler(scheduler)
# d:0.0088 o:0.0042, c:0.0098
# steps_per_epoch=train_num,
# validation_steps=valid_num,
# callbacks=[early_stop, checkpoint, tensorboard],
model.fit_generator(generator=train_gen,
steps_per_epoch=train_num,
epochs=500,
verbose=1,
validation_data=valid_gen,
validation_steps=valid_num,
shuffle=True,
callbacks=[checkpoint, tensorboard, lr_sched],
max_queue_size=3)
def predict(args):
# complie models
model = nn.network()
# model.load_weights('3dbox_weights_1st.hdf5')
model.load_weights(args.w)
# KITTI_train_gen = KITTILoader(subset='training')
dims_avg, _ = KITTILoader(subset='training').get_average_dimension()
print("dims_avg = ", dims_avg)
# dims_avg = {'Car': array([1.52608343, 1.62858987, 3.88395449])}
# list all the validation images
if args.a == 'training':
all_imgs = sorted(os.listdir(test_image_dir))
val_index = int(len(all_imgs) * cfg().split)
val_imgs = all_imgs[val_index:]
else:
val_imgs = sorted(os.listdir(test_image_dir))
start_time = time.time()
for i in val_imgs:
image_file = test_image_dir + i
depth_file = test_depth_dir + i
label_file = test_label_dir + i.replace('png', 'txt')
prediction_file = prediction_path + i.replace('png', 'txt')
calibration_file = test_calib_path + i.replace('png', 'txt')
#calibration_file = os.path.join('/media/ferdyan/NewDisk/Trajectory_Final/bbox_3d/0000.txt')
# write the prediction file
with open(prediction_file, 'w') as predict:
img = cv2.imread(image_file)
img = np.array(img, dtype='float32')
dpth = cv2.imread(depth_file)
dpth = np.array(dpth, dtype='float32')
P2 = np.array([])
for line in open(calibration_file):
if 'P2' in line:
P2 = line.split(' ')
P2 = np.asarray([float(i) for i in P2[1:]])
P2 = np.reshape(P2, (3, 4))
for line in open(label_file):
line = line.strip().split(' ')
#print("line = ", line)
obj = detectionInfo(line)
xmin = int(obj.xmin)
xmax = int(obj.xmax)
ymin = int(obj.ymin)
ymax = int(obj.ymax)
box2d = [xmin, ymin, xmax, ymax]
box_2D = np.asarray(box2d, dtype=np.float)
if obj.name in cfg().KITTI_cat:
# cropped 2d bounding box
if xmin == xmax or ymin == ymax:
continue
# 2D detection area RGB image
patch = img[ymin:ymax, xmin:xmax]
patch = cv2.resize(patch, (cfg().norm_h, cfg().norm_w))
patch -= np.array([[[103.939, 116.779, 123.68]]])
# extend it to match the training dimension
patch = np.expand_dims(patch, 0)
# 2D detection area depth map
#patch_d = dpth[ymin : ymax, xmin : xmax]
#patch_d = cv2.resize(patch_d, (cfg().norm_h, cfg().norm_w))
#patch_d -= np.array([[[103.939, 116.779, 123.68]]])
# extend it to match the training dimension
#patch_d = np.expand_dims(patch_d, 0)
# one
prediction = model.predict([patch])
# two
#prediction = model.predict([patch, patch_d])
# TAMBAHAN AING
# Transform regressed angle
box2d_center_x = (xmin + xmax) / 2.0
theta_ray = np.arctan(fx / (box2d_center_x - u0))
if theta_ray < 0:
theta_ray = theta_ray + np.pi
max_anc = np.argmax(prediction[2][0])
anchors = prediction[1][0][max_anc]
if anchors[1] > 0:
angle_offset = np.arccos(anchors[0])
else:
angle_offset = -np.arccos(anchors[0])
bin_num = prediction[2][0].shape[0]
wedge = 2. * np.pi / bin_num
theta_loc = angle_offset + max_anc * wedge
theta = theta_loc + theta_ray
# object's yaw angle
yaw = np.pi / 2 - theta
points2D = gen_3D_box(yaw, dims, cam_to_img, box_2D)
draw_3D_box(img, points2D)
cv2.imshow('f', img)
cv2.waitKey(0)
cv2.destroyAllWindows()
#cv2.imwrite('output/'+ f.replace('png','jpg'), img)
end_time = time.time()
process_time = (end_time - start_time) / len(val_imgs)
print(process_time)
def train():
KITTI_train_gen = KITTILoader(subset='training')
dim_avg, dim_cnt = KITTI_train_gen.get_average_dimension()
new_data = orientation_confidence_flip(KITTI_train_gen.image_data, dim_avg)
model = nn.network()
# model.load_weights('3dbox_weights_mob.hdf5')
early_stop = callbacks.EarlyStopping(monitor='val_loss',
min_delta=0.001,
patience=10,
mode='min',
verbose=1)
checkpoint = callbacks.ModelCheckpoint('3dbox_weights_mob.hdf5',
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min',
period=1)
tensorboard = callbacks.TensorBoard(log_dir='logs/',
histogram_freq=0,
write_graph=True,
write_images=False)
all_examples = len(new_data)
trv_split = int(cfg().split * all_examples) # train val split
train_gen = data_gen(new_data[:trv_split])
valid_gen = data_gen(new_data[trv_split:all_examples])
train_num = int(np.ceil(trv_split / cfg().batch_size))
valid_num = int(np.ceil((all_examples - trv_split) / cfg().batch_size))
# choose the minimizer to be sgd
minimizer = optimizer.SGD(lr=0.0001, momentum=0.9)
# multi task learning
model.compile(
optimizer=minimizer, #minimizer,
loss={
'dimensions': 'mean_squared_error',
'orientation': orientation_loss,
'confidence': 'binary_crossentropy'
},
loss_weights={
'dimensions': 1.,
'orientation': 10.,
'confidence': 5.
})
# d:0.0088 o:0.0042, c:0.0098
model.fit_generator(generator=train_gen,
steps_per_epoch=train_num,
epochs=500,
verbose=1,
validation_data=valid_gen,
validation_steps=valid_num,
shuffle=True,
callbacks=[early_stop, checkpoint, tensorboard],
max_queue_size=3)
def train():
KITTI_train_gen = KITTILoader(subset='training')
dim_avg, dim_cnt = KITTI_train_gen.get_average_dimension()
new_data = orientation_confidence_flip(KITTI_train_gen.image_data, dim_avg)
model = nn.network()
# model.load_weights('3dbox_weights_mob.hdf5')
early_stop = callbacks.EarlyStopping(monitor='val_loss',
min_delta=0.001,
patience=10,
mode='min',
verbose=1)
checkpoint = callbacks.ModelCheckpoint(
'3dbox_mbnv2_{}x{}_float32.hdf5'.format(cfg().norm_h,
cfg().norm_w),
monitor='val_loss',
verbose=1,
save_best_only=True,
mode='min',
period=1)
tensorboard = callbacks.TensorBoard(log_dir='logs/',
histogram_freq=0,
write_graph=True,
write_images=False)
all_examples = len(new_data)
trv_split = int(cfg().split * all_examples) # train val split
train_gen = data_gen(new_data[:trv_split])
valid_gen = data_gen(new_data[trv_split:all_examples])
train_num = int(np.ceil(trv_split / cfg().batch_size))
valid_num = int(np.ceil((all_examples - trv_split) / cfg().batch_size))
# choose the minimizer to be sgd
minimizer = optimizer.SGD(lr=0.0001, momentum=0.9)
# multi task learning
model.compile(
optimizer=minimizer, #minimizer,
loss={
'dimensions': 'mean_squared_error',
'orientation': orientation_loss,
'confidence': 'binary_crossentropy'
},
loss_weights={
'dimensions': 1.,
'orientation': 10.,
'confidence': 5.
})
# d:0.0088 o:0.0042, c:0.0098
model.fit_generator(generator=train_gen,
steps_per_epoch=train_num,
epochs=500,
verbose=1,
validation_data=valid_gen,
validation_steps=valid_num,
shuffle=True,
callbacks=[early_stop, checkpoint, tensorboard],
max_queue_size=3)
tf.saved_model.save(model,
'saved_model_{}x{}'.format(cfg().norm_h,
cfg().norm_w))
model.save('3dbox_mbnv2_{}x{}_float32.h5'.format(cfg().norm_h,
cfg().norm_w))
full_model = tf.function(lambda inputs: model(inputs))
full_model = full_model.get_concrete_function(
inputs=(tf.TensorSpec(model.inputs[0].shape, model.inputs[0].dtype)))
frozen_func = convert_variables_to_constants_v2(full_model,
lower_control_flow=False)
frozen_func.graph.as_graph_def()
tf.io.write_graph(graph_or_graph_def=frozen_func.graph,
logdir=".",
name="3dbox_mbnv2_{}x{}_float32.pb".format(
cfg().norm_h,
cfg().norm_w),
as_text=False)
import time
import math
from config import config as cfg
if cfg().network == 'vgg16':
from model import vgg16 as nn
if cfg().network == 'mobilenet_v2':
from model import mobilenet_v2_early_element as nn
if cfg().network == 'vgg16v2':
from model import vgg16v2 as nn
if cfg().network == 'vgg16_one':
from model import vgg16_minggu_pre as nn
# Construct the network
model = nn.network()
model.load_weights(
r'/media/ferdyan/NewDisk/3d_bbox_depth/Weights_train/weight_7Okt_Sore/model00000125.hdf5'
)
image_dir = '/media/ferdyan/NewDisk/3d_bounding_box/train_dataset/2011_09_26/own/image_02/data/'
calib_dir = '/media/ferdyan/NewDisk/3d_bounding_box/train_dataset/2011_09_26/own/calib_02/'
box2d_dir = '/media/ferdyan/NewDisk/3d_bounding_box/train_dataset/2011_09_26/own/label_02/'
classes = ['Car']
cls_to_ind = {cls: i for i, cls in enumerate(classes)}
dims_avg = np.loadtxt(r'/media/ferdyan/NewDisk/3d_bbox_depth/voc_dims.txt',
delimiter=',')
