def localise_all(model, img_paths):
x, q = [], []
std_x, std_q = [], []
with Localiser(model,
uncertainty=False,
output_type='axis',
dropout=args.dropout) as localiser:
for i, path in enumerate(img_paths):
img = read_image(path, expand_dims=False, normalise=True)
img = resize_image(centre_crop(img), args.size)
pred = localiser.localise(img, samples=args.samples)
x.append(pred['x'])
q.append(pred['q'])
#std_x.append(pred['std_x_all'])
#std_q.append(pred['std_q_all'])
return np.array(x), np.array(q) #, np.array(std_x), np.array(std_q)
else:
imgs = []
n_images = len(imgs)
if n_images > 1 and not args.output:
print('--output argument required')
sys.exit(1)
if n_images == 0:
print('No images found')
sys.exit(1)
if args.output and not os.path.isdir(args.output):
os.makedirs(args.output)
# Localise
with Localiser(args.model,
output_type='axis' if args.axis else 'quat') as localiser:
for i in range(n_images):
# Load normalised image
image = read_image(imgs[i], normalise=True)
image = resize_image(centre_crop(image), args.size)
# Compute the saliency map
grad = localiser.saliency(image)
grad = median(grad, disk(3))
if args.scale:
grad = np.clip(1.0 * grad / args.scale, 0, 1)
else:
grad = 1.0 * grad / np.max(grad)
if args.red:
grad = np.stack([
if args.model:
input_size = 256
test_reader = ImageReader(args.dataset,
batch_size=1,
image_size=[input_size, input_size],
randomise=False)
n_images = test_reader.total_images()
positions = np.empty([0, 3])
orientations = np.empty([0, 4])
if args.uncertainty:
std_x = []
std_q = []
with Localiser(args.model, uncertainty=args.uncertainty) as localiser:
for i in range(n_images):
images_feed, labels_feed = test_reader.next_batch()
# Make prediction
predicted = localiser.localise(images_feed)
positions = np.concatenate(
(positions, np.asarray([predicted['x']])))
orientations = np.concatenate(
(orientations, np.asarray([predicted['q']])))
if args.uncertainty:
std_x.append(predicted['std_x'])
std_q.append(predicted['std_q'])
progress_bar(1.0 * (i + 1) / n_images, 30, text='Localising')
print('')
n_images = len(imgs)
if n_images > 1 and not args.output:
print('--output argument required')
sys.exit(1)
if n_images == 0:
print('No images found')
sys.exit(1)
if args.output and not os.path.isdir(args.output):
os.makedirs(args.output)
# Localise
input_size = 256
with Localiser(args.model) as localiser:
for i in range(n_images):
# Load normalised image
image = read_image(imgs[i], normalise=True, size=[input_size, input_size])
# Compute the saliency map
grad = localiser.saliency(image)
grad = median(grad, disk(3))
grad = 1.0*grad/np.max(grad)
if args.output:
fname = os.path.join(args.output, os.path.basename(imgs[i]))
scipy.misc.imsave(fname, grad)
progress_bar(1.0*(i+1)/n_images, 30, text='Computing')
else:
# Display image
import matplotlib.pyplot as plt
input_size = 224
test_reader = ImageReader(args.dataset,
batch_size=1,
image_size=[input_size, input_size],
random_crop=False,
randomise=False)
n_images = test_reader.total_images()
pos_errors = []
orient_errors = []
positions = np.empty([0, 3])
orientations = np.empty([0, 4])
with Localiser(input_size, args.model) as localiser:
for i in range(n_images):
images_feed, labels_feed = test_reader.next_batch()
gt = {'x': labels_feed[0][0:3], 'q': labels_feed[0][3:7]}
# Make prediction
predicted = localiser.localise(images_feed)
x, y, z = predicted['x']
q1, q2, q3, q4 = predicted['q']
pos_error = l2_distance(gt['x'], predicted['x'])
orient_error = quaternion_distance(gt['q'],
predicted['q']) * 180 / np.pi
pos_errors.append(pos_error)
orient_errors.append(orient_error)
#x_min, x_max = 0.5, 5.997787
img_paths, labels = read_label_file(dataset,
full_paths=True,
convert_to_axis=True)
n_images = len(img_paths)
x_scale = np.linspace(x_min, x_max, n_images)
x_gt = map(lambda l: l[0:3], labels)
q_gt = map(lambda l: l[3:], labels)
x = []
q = []
std_x = []
std_q = []
with Localiser(model, uncertainty=False, output_type='axis',
dropout=0.5) as localiser:
for i, path in enumerate(img_paths):
img = read_image(path, expand_dims=False, normalise=True)
img = resize_image(centre_crop(img), [input_size, input_size])
pred = localiser.localise(img, samples=40)
x.append(pred['x'])
q.append(pred['q'])
if 'std_x_all' in pred:
std_x.append(pred['std_x_all'])
std_q.append(pred['std_q_all'])
progress_bar(1.0 * (i + 1) / n_images, 30, text='Localising')
print ''
x_gt = np.array(x_gt)
q_gt = np.array(q_gt)
x = np.array(x)
args = parser.parse_args()
input_size = 224
test_reader = ImageReader(args.dataset,
batch_size=1,
image_size=[input_size, input_size],
random_crop=False,
randomise=False)
n_images = test_reader.total_images()
# Read the definition file to get file names
paths, _ = read_label_file(args.dataset)
paths = map(lambda x: os.path.basename(x), paths)
# Localise
with Localiser(input_size, args.model,
uncertainty=args.uncertainty) as localiser:
if args.output:
f = open(args.output, 'w')
f.write('\n\n\n') # Hack for now
for i in range(n_images):
images_feed, labels_feed = test_reader.next_batch()
# Make prediction
predicted = localiser.localise(images_feed)
x = [round(v, 6) for v in predicted['x']]
q = [round(v, 6) for v in predicted['q']]
if args.output:
f.write('{} {} {} {} {} {} {} {}\n'.format(paths[i], x[0], x[1],
x[2], q[0], q[1], q[2],
test_reader = ImageReader(args.dataset,
batch_size=1,
image_size=input_size,
crop_size=crop_size,
centre_crop=True,
randomise=False)
n_images = test_reader.total_images()
positions = []
orientations = []
if args.uncertainty:
std_x = []
std_q = []
with Localiser(args.model,
uncertainty=args.uncertainty,
output_type=output_type,
dropout=0.3) as localiser:
for i in range(n_images):
images_feed, labels_feed = test_reader.next_batch()
# Make prediction
predicted = localiser.localise(images_feed)
positions.append(predicted['x'])
orientations.append(predicted['q'])
if args.uncertainty:
std_x.append(predicted['std_x'])
std_q.append(predicted['std_q'])
progress_bar(1.0 * (i + 1) / n_images, 30, text='Localising')
print('')
positions = np.array(positions)