def load_and_transform3d(ex, joint_info, learning_phase, rng):
# Get the random number generators for the different augmentations to make it reproducibile
appearance_rng = util.new_rng(rng)
background_rng = util.new_rng(rng)
geom_rng = util.new_rng(rng)
partial_visi_rng = util.new_rng(rng)
output_side = FLAGS.proc_side
output_imshape = (output_side, output_side)
if 'sailvos' in ex.image_path.lower():
# This is needed in order not to lose precision in later operations.
# Background: In the Sailvos dataset (GTA V), some world coordinates
# are crazy large (several kilometers, i.e. millions of millimeters, which becomes
# hard to process with the limited simultaneous dynamic range of float32).
# They are stored in float64 but the processing is done in float32 here.
ex.world_coords -= ex.camera.t
ex.camera.t[:] = 0
box = ex.bbox
if 'surreal' in ex.image_path.lower():
# Surreal images are flipped wrong in the official dataset release
box = box.copy()
box[0] = 320 - (box[0] + box[2])
# Partial visibility
if 'surreal' in ex.image_path.lower() and 'surmuco' not in FLAGS.dataset:
partial_visi_prob = 0.5
elif 'h36m' in ex.image_path.lower() and 'many' in FLAGS.dataset:
partial_visi_prob = 0.5
else:
partial_visi_prob = FLAGS.partial_visibility_prob
use_partial_visi_aug = ((learning_phase == TRAIN or FLAGS.test_aug)
and partial_visi_rng.rand() < partial_visi_prob)
if use_partial_visi_aug:
box = util.random_partial_subbox(boxlib.expand_to_square(box),
partial_visi_rng)
# Geometric transformation and augmentation
crop_side = np.max(box[2:])
center_point = boxlib.center(box)
if ((learning_phase == TRAIN and FLAGS.geom_aug) or
(learning_phase != TRAIN and FLAGS.test_aug and FLAGS.geom_aug)):
center_point += util.random_uniform_disc(
geom_rng) * FLAGS.shift_aug / 100 * crop_side
# The homographic reprojection of a rectangle (bounding box) will not be another rectangle
# Hence, instead we transform the side midpoints of the short sides of the box and
# determine an appropriate zoom factor by taking the projected distance of these two points
# and scaling that to the desired output image side length.
if box[2] < box[3]:
# Tall box: take midpoints of top and bottom sides
delta_y = np.array([0, box[3] / 2])
sidepoints = center_point + np.stack([-delta_y, delta_y])
else:
# Wide box: take midpoints of left and right sides
delta_x = np.array([box[2] / 2, 0])
sidepoints = center_point + np.stack([-delta_x, delta_x])
cam = ex.camera.copy()
cam.turn_towards(target_image_point=center_point)
cam.undistort()
cam.square_pixels()
cam_sidepoints = cameralib.reproject_image_points(sidepoints, ex.camera,
cam)
crop_side = np.linalg.norm(cam_sidepoints[0] - cam_sidepoints[1])
cam.zoom(output_side / crop_side)
cam.center_principal_point(output_imshape)
if FLAGS.geom_aug and (learning_phase == TRAIN or FLAGS.test_aug):
s1 = FLAGS.scale_aug_down / 100
s2 = FLAGS.scale_aug_up / 100
zoom = geom_rng.uniform(1 - s1, 1 + s2)
cam.zoom(zoom)
r = np.deg2rad(FLAGS.rot_aug)
cam.rotate(roll=geom_rng.uniform(-r, r))
world_coords = ex.univ_coords if FLAGS.universal_skeleton else ex.world_coords
metric_world_coords = ex.world_coords
if learning_phase == TRAIN and geom_rng.rand() < 0.5:
cam.horizontal_flip()
# Must reorder the joints due to left and right flip
camcoords = cam.world_to_camera(world_coords)[
joint_info.mirror_mapping]
metric_world_coords = metric_world_coords[joint_info.mirror_mapping]
else:
camcoords = cam.world_to_camera(world_coords)
imcoords = cam.world_to_image(metric_world_coords)
# Load and reproject image
image_path = util.ensure_absolute_path(ex.image_path)
origsize_im = improc.imread_jpeg(image_path)
if 'surreal' in ex.image_path.lower():
# Surreal images are flipped wrong in the official dataset release
origsize_im = origsize_im[:, ::-1]
interp_str = (FLAGS.image_interpolation_train if learning_phase == TRAIN
else FLAGS.image_interpolation_test)
antialias = (FLAGS.antialias_train
if learning_phase == TRAIN else FLAGS.antialias_test)
interp = getattr(cv2, 'INTER_' + interp_str.upper())
im = cameralib.reproject_image(origsize_im,
ex.camera,
cam,
output_imshape,
antialias_factor=antialias,
interp=interp)
# Color adjustment
if re.match('.*mupots/TS[1-5]/.+', ex.image_path):
im = improc.adjust_gamma(im, 0.67, inplace=True)
elif '3dhp' in ex.image_path and re.match('.+/(TS[1-4])/', ex.image_path):
im = improc.adjust_gamma(im, 0.67, inplace=True)
im = improc.white_balance(im, 110, 145)
elif 'panoptic' in ex.image_path.lower():
im = improc.white_balance(im, 120, 138)
# Background augmentation
if hasattr(ex, 'mask') and ex.mask is not None:
bg_aug_prob = 0.2 if 'sailvos' in ex.image_path.lower(
) else FLAGS.background_aug_prob
if (FLAGS.background_aug_prob
and (learning_phase == TRAIN or FLAGS.test_aug)
and background_rng.rand() < bg_aug_prob):
fgmask = improc.decode_mask(ex.mask)
if 'surreal' in ex.image_path:
# Surreal images are flipped wrong in the official dataset release
fgmask = fgmask[:, ::-1]
fgmask = cameralib.reproject_image(fgmask,
ex.camera,
cam,
output_imshape,
antialias_factor=antialias,
interp=interp)
im = augmentation.background.augment_background(
im, fgmask, background_rng)
# Occlusion and color augmentation
im = augmentation.appearance.augment_appearance(im, learning_phase,
FLAGS.occlude_aug_prob,
appearance_rng)
im = tfu.nhwc_to_std(im)
im = improc.normalize01(im)
# Joints with NaN coordinates are invalid
is_joint_in_fov = ~np.logical_or(
np.any(imcoords < 0, axis=-1),
np.any(imcoords >= FLAGS.proc_side, axis=-1))
joint_validity_mask = ~np.any(np.isnan(camcoords), axis=-1)
rot_to_orig_cam = ex.camera.R @ cam.R.T
rot_to_world = cam.R.T
return dict(image=im,
intrinsics=np.float32(cam.intrinsic_matrix),
image_path=ex.image_path,
coords3d_true=np.nan_to_num(camcoords).astype(np.float32),
coords2d_true=np.nan_to_num(imcoords).astype(np.float32),
rot_to_orig_cam=rot_to_orig_cam.astype(np.float32),
rot_to_world=rot_to_world.astype(np.float32),
cam_loc=cam.t.astype(np.float32),
joint_validity_mask=joint_validity_mask,
is_joint_in_fov=np.float32(is_joint_in_fov))
def load_and_transform3d(ex, joint_info, learning_phase, rng=None):
appearance_rng = util.new_rng(rng)
background_rng = util.new_rng(rng)
geom_rng = util.new_rng(rng)
partial_visi_rng = util.new_rng(rng)
output_side = FLAGS.proc_side
output_imshape = (output_side, output_side)
box = ex.bbox
if FLAGS.partial_visibility:
box = util.random_partial_subbox(boxlib.expand_to_square(box), partial_visi_rng)
crop_side = np.max(box[2:])
center_point = boxlib.center(box)
if ((learning_phase == TRAIN and FLAGS.geom_aug) or
(learning_phase != TRAIN and FLAGS.test_aug and FLAGS.geom_aug)):
center_point += util.random_uniform_disc(geom_rng) * FLAGS.shift_aug / 100 * crop_side
if box[2] < box[3]:
delta_y = np.array([0, box[3] / 2])
sidepoints = center_point + np.stack([-delta_y, delta_y])
else:
delta_x = np.array([box[2] / 2, 0])
sidepoints = center_point + np.stack([-delta_x, delta_x])
cam = ex.camera.copy()
cam.turn_towards(target_image_point=center_point)
cam.undistort()
cam.square_pixels()
world_sidepoints = ex.camera.image_to_world(sidepoints)
cam_sidepoints = cam.world_to_image(world_sidepoints)
crop_side = np.linalg.norm(cam_sidepoints[0] - cam_sidepoints[1])
cam.zoom(output_side / crop_side)
cam.center_principal_point(output_imshape)
if FLAGS.geom_aug and (learning_phase == TRAIN or FLAGS.test_aug):
s1 = FLAGS.scale_aug_down / 100
s2 = FLAGS.scale_aug_up / 100
r = FLAGS.rot_aug * np.pi / 180
zoom = geom_rng.uniform(1 - s1, 1 + s2)
cam.zoom(zoom)
cam.rotate(roll=geom_rng.uniform(-r, r))
world_coords = ex.univ_coords if FLAGS.universal_skeleton else ex.world_coords
metric_world_coords = ex.world_coords
if learning_phase == TRAIN and geom_rng.rand() < 0.5:
cam.horizontal_flip()
camcoords = cam.world_to_camera(world_coords)[joint_info.mirror_mapping]
metric_world_coords = metric_world_coords[joint_info.mirror_mapping]
else:
camcoords = cam.world_to_camera(world_coords)
imcoords = cam.world_to_image(metric_world_coords)
image_path = util.ensure_absolute_path(ex.image_path)
origsize_im = improc.imread_jpeg(image_path)
interp_str = (FLAGS.image_interpolation_train
if learning_phase == TRAIN else FLAGS.image_interpolation_test)
antialias = (FLAGS.antialias_train if learning_phase == TRAIN else FLAGS.antialias_test)
interp = getattr(cv2, 'INTER_' + interp_str.upper())
im = cameralib.reproject_image(
origsize_im, ex.camera, cam, output_imshape, antialias_factor=antialias, interp=interp)
if re.match('.+/mupots/TS[1-5]/.+', ex.image_path):
im = improc.adjust_gamma(im, 0.67, inplace=True)
elif '3dhp' in ex.image_path and re.match('.+/(TS[1-4])/', ex.image_path):
im = improc.adjust_gamma(im, 0.67, inplace=True)
im = improc.white_balance(im, 110, 145)
if (FLAGS.background_aug_prob and hasattr(ex, 'mask') and ex.mask is not None and
background_rng.rand() < FLAGS.background_aug_prob and
(learning_phase == TRAIN or FLAGS.test_aug)):
fgmask = improc.decode_mask(ex.mask)
fgmask = cameralib.reproject_image(
fgmask, ex.camera, cam, output_imshape, antialias_factor=antialias, interp=interp)
im = augmentation.background.augment_background(im, fgmask, background_rng)
im = augmentation.appearance.augment_appearance(im, learning_phase, appearance_rng)
im = tfu.nhwc_to_std(im)
im = improc.normalize01(im)
# Joints with NaN coordinates are invalid
is_joint_in_fov = ~np.logical_or(np.any(imcoords < 0, axis=-1),
np.any(imcoords >= FLAGS.proc_side, axis=-1))
joint_validity_mask = ~np.any(np.isnan(camcoords), axis=-1)
rot_to_orig_cam = ex.camera.R @ cam.R.T
rot_to_world = cam.R.T
inv_intrinsics = np.linalg.inv(cam.intrinsic_matrix)
return (
ex.image_path, im, np.nan_to_num(camcoords).astype(np.float32),
np.nan_to_num(imcoords).astype(np.float32), inv_intrinsics.astype(np.float32),
rot_to_orig_cam.astype(np.float32), rot_to_world.astype(np.float32),
cam.t.astype(np.float32), joint_validity_mask,
np.float32(is_joint_in_fov), ex.activity_name, ex.scene_name)
def make_efficient_example(ex,
new_image_path,
further_expansion_factor=1,
image_adjustments_3dhp=False,
min_time=None):
"""Make example by storing the image in a cropped and resized version for efficient loading"""
is3d = hasattr(ex, 'world_coords')
w, h = (improc.image_extents(util.ensure_absolute_path(ex.image_path))
if isinstance(ex.image_path, str) else
(ex.image_path.shape[1], ex.image_path.shape[0]))
full_box = boxlib.full_box(imsize=[w, h])
if is3d:
old_camera = ex.camera
new_camera = ex.camera.copy()
new_camera.turn_towards(target_image_point=boxlib.center(ex.bbox))
new_camera.undistort()
else:
old_camera = cameralib.Camera.create2D()
new_camera = old_camera.copy()
reprojected_box = reproject_box(ex.bbox,
old_camera,
new_camera,
method='side_midpoints')
reprojected_full_box = reproject_box(full_box,
old_camera,
new_camera,
method='corners')
expanded_bbox = (get_expanded_crop_box(
reprojected_box, reprojected_full_box, further_expansion_factor)
if further_expansion_factor > 0 else reprojected_box)
scale_factor = min(1.2, 256 / np.max(reprojected_box[2:]) * 1.5)
new_camera.shift_image(-expanded_bbox[:2])
new_camera.scale_output(scale_factor)
reprojected_box = reproject_box(ex.bbox,
old_camera,
new_camera,
method='side_midpoints')
dst_shape = improc.rounded_int_tuple(scale_factor * expanded_bbox[[3, 2]])
new_image_abspath = util.ensure_absolute_path(new_image_path)
if not (util.is_file_newer(new_image_abspath, min_time)
and improc.is_image_readable(new_image_abspath)):
im = improc.imread_jpeg(ex.image_path) if isinstance(
ex.image_path, str) else ex.image_path
#host_im, cuda_im = get_memory(im.shape)
im = np.power((im.astype(np.float32) / 255), 2.2)
#cuda_im.upload(host_im)
new_im = cameralib.reproject_image(im,
old_camera,
new_camera,
dst_shape,
antialias_factor=2,
interp=cv2.INTER_CUBIC)
new_im = np.clip(new_im, 0, 1)
if image_adjustments_3dhp:
# enhance the 3dhp images to reduce the green tint and increase brightness
new_im = (new_im**(1 / 2.2 * 0.67) * 255).astype(np.uint8)
new_im = improc.white_balance(new_im, 110, 145)
else:
new_im = (new_im**(1 / 2.2) * 255).astype(np.uint8)
util.ensure_path_exists(new_image_abspath)
imageio.imwrite(new_image_abspath, new_im, quality=95)
assert improc.is_image_readable(new_image_abspath)
new_ex = copy.deepcopy(ex)
new_ex.bbox = reprojected_box
new_ex.image_path = new_image_path
if is3d:
new_ex.camera = new_camera
else:
new_ex.coords = cameralib.reproject_image_points(
new_ex.coords, old_camera, new_camera)
if hasattr(ex, 'mask') and ex.mask is not None:
if isinstance(ex.mask, str):
mask = improc.imread_jpeg(util.ensure_absolute_path(ex.mask))
host_mask, cuda_mask = get_memory(mask.shape)
np.divide(mask.astype(np.float32), 255, out=host_mask)
cuda_mask.upload(host_mask)
mask_reproj = cameralib.reproject_image(
cuda_mask,
ex.camera,
new_camera,
dst_shape,
antialias_factor=2).download()
mask_reproj = 255 * (mask_reproj[..., 0] > 32 / 255).astype(
np.uint8)
new_ex.mask = get_connected_component_with_highest_iou(
mask_reproj, reprojected_box)
else:
new_ex.mask = ex.mask
return new_ex