def train_transform(rgb, sparse, target, rgb_near, args):
# s = np.random.uniform(1.0, 1.5) # random scaling
# angle = np.random.uniform(-5.0, 5.0) # random rotation degrees
do_flip = np.random.uniform(0.0, 1.0) < 0.5 # random horizontal flip
transform_geometric = transforms.Compose([
# transforms.Rotate(angle),
# transforms.Resize(s),
transforms.BottomCrop((oheight, owidth)),
transforms.HorizontalFlip(do_flip)
])
if sparse is not None:
sparse = transform_geometric(sparse)
target = transform_geometric(target)
if rgb is not None:
brightness = np.random.uniform(max(0, 1 - args.jitter),
1 + args.jitter)
contrast = np.random.uniform(max(0, 1 - args.jitter), 1 + args.jitter)
saturation = np.random.uniform(max(0, 1 - args.jitter),
1 + args.jitter)
transform_rgb = transforms.Compose([
transforms.ColorJitter(brightness, contrast, saturation, 0),
transform_geometric
])
rgb = transform_rgb(rgb)
if rgb_near is not None:
rgb_near = transform_rgb(rgb_near)
# sparse = drop_depth_measurements(sparse, 0.9)
return rgb, sparse, target, rgb_near
def val_transform(rgb, sparse, target, rgb_near, args):
transform = transforms.Compose([
transforms.BottomCrop((oheight, owidth)),
])
if rgb is not None:
rgb = transform(rgb)
if sparse is not None:
sparse = transform(sparse)
if target is not None:
target = transform(target)
if rgb_near is not None:
rgb_near = transform(rgb_near)
return rgb, sparse, target, rgb_near
def val_transform(rgb, sparse, target, position, args):
oheight = args.val_h
owidth = args.val_w
transform = transforms.Compose([
transforms.BottomCrop((oheight, owidth)),
])
if rgb is not None:
rgb = transform(rgb)
if sparse is not None:
sparse = transform(sparse)
if target is not None:
target = transform(target)
if position is not None:
position = transform(position)
return rgb, sparse, target, position
def train_transform(rgb, sparse, target, position, args):
# s = np.random.uniform(1.0, 1.5) # random scaling
# angle = np.random.uniform(-5.0, 5.0) # random rotation degrees
oheight = args.val_h
owidth = args.val_w
do_flip = np.random.uniform(0.0, 1.0) < 0.5 # random horizontal flip
transforms_list = [
# transforms.Rotate(angle),
# transforms.Resize(s),
transforms.BottomCrop((oheight, owidth)),
transforms.HorizontalFlip(do_flip)
]
# if small_training == True:
# transforms_list.append(transforms.RandomCrop((rheight, rwidth)))
transform_geometric = transforms.Compose(transforms_list)
if sparse is not None:
sparse = transform_geometric(sparse)
target = transform_geometric(target)
if rgb is not None:
brightness = np.random.uniform(max(0, 1 - args.jitter),
1 + args.jitter)
contrast = np.random.uniform(max(0, 1 - args.jitter), 1 + args.jitter)
saturation = np.random.uniform(max(0, 1 - args.jitter),
1 + args.jitter)
transform_rgb = transforms.Compose([
transforms.ColorJitter(brightness, contrast, saturation, 0),
transform_geometric
])
rgb = transform_rgb(rgb)
# sparse = drop_depth_measurements(sparse, 0.9)
if position is not None:
bottom_crop_only = transforms.Compose(
[transforms.BottomCrop((oheight, owidth))])
position = bottom_crop_only(position)
# random crop
#if small_training == True:
if args.not_random_crop == False:
h = oheight
w = owidth
rheight = args.random_crop_height
rwidth = args.random_crop_width
# randomlize
i = np.random.randint(0, h - rheight + 1)
j = np.random.randint(0, w - rwidth + 1)
if rgb is not None:
if rgb.ndim == 3:
rgb = rgb[i:i + rheight, j:j + rwidth, :]
elif rgb.ndim == 2:
rgb = rgb[i:i + rheight, j:j + rwidth]
if sparse is not None:
if sparse.ndim == 3:
sparse = sparse[i:i + rheight, j:j + rwidth, :]
elif sparse.ndim == 2:
sparse = sparse[i:i + rheight, j:j + rwidth]
if target is not None:
if target.ndim == 3:
target = target[i:i + rheight, j:j + rwidth, :]
elif target.ndim == 2:
target = target[i:i + rheight, j:j + rwidth]
if position is not None:
if position.ndim == 3:
position = position[i:i + rheight, j:j + rwidth, :]
elif position.ndim == 2:
position = position[i:i + rheight, j:j + rwidth]
return rgb, sparse, target, position
