in visualization/POF/data/BaseReader.py [0:0]
def get(self, withPAF=True, PAF_normalize3d=True, read_image=True, imw=1920, imh=1080, bbox2d=0):
# bbox2d: 0: computed from 3D bounding box, 1: compute from openpose
assert bbox2d in (0, 1)
assert type(withPAF) == bool
# produce data from slice_input_producer
flow_list = tf.train.slice_input_producer(list(self.tensor_dict.values()), shuffle=self.shuffle)
flow_dict = {key: flow_list[ik] for ik, key in enumerate(self.tensor_dict.keys())}
# build data dictionary
data_dict = {}
data_dict['img_dir'] = flow_dict['img_dirs']
data_dict['K'] = flow_dict['K']
# rotate and project to camera frame
if self.objtype == 0:
body2d, body3d = self.project_tf(flow_dict['body'], flow_dict['K'], flow_dict['R'], flow_dict['t'], flow_dict['distCoef'])
body3d = tf.cast(body3d, tf.float32)
body2d = tf.cast(body2d, tf.float32)
data_dict['keypoint_xyz_origin'] = body3d
data_dict['keypoint_uv_origin'] = body2d
data_dict['body_valid'] = flow_dict['body_valid']
elif self.objtype == 1:
cond_left = tf.reduce_any(tf.cast(flow_dict['left_hand_valid'], dtype=tf.bool)) # 0 for right hand, 1 for left hand
hand3d = tf.cond(cond_left, lambda: flow_dict['left_hand'], lambda: flow_dict['right_hand']) # in world coordinate
hand2d, hand3d = self.project_tf(hand3d, flow_dict['K'], flow_dict['R'], flow_dict['t'], flow_dict['distCoef']) # in camera coordinate
hand3d = tf.cast(hand3d, tf.float32)
hand2d = tf.cast(hand2d, tf.float32)
data_dict['keypoint_xyz_origin'] = hand3d
data_dict['keypoint_uv_origin'] = hand2d
data_dict['cond_left'] = cond_left
data_dict['left_hand_valid'] = flow_dict['left_hand_valid']
data_dict['right_hand_valid'] = flow_dict['right_hand_valid']
elif self.objtype == 2:
body2d, body3d = self.project_tf(flow_dict['body'], flow_dict['K'], flow_dict['R'], flow_dict['t'], flow_dict['distCoef'])
lhand2d, lhand3d = self.project_tf(flow_dict['left_hand'], flow_dict['K'], flow_dict['R'], flow_dict['t'], flow_dict['distCoef'])
rhand2d, rhand3d = self.project_tf(flow_dict['right_hand'], flow_dict['K'], flow_dict['R'], flow_dict['t'], flow_dict['distCoef'])
data_dict['body_xyz_origin'] = body3d
data_dict['body_uv_origin'] = body2d
data_dict['lhand_xyz_origin'] = lhand3d
data_dict['lhand_uv_origin'] = lhand2d
data_dict['rhand_xyz_origin'] = rhand3d
data_dict['rhand_uv_origin'] = rhand2d
data_dict['body_valid'] = flow_dict['body_valid']
data_dict['left_hand_valid'] = flow_dict['left_hand_valid']
data_dict['right_hand_valid'] = flow_dict['right_hand_valid']
# read image
if read_image:
img_file = tf.read_file(flow_dict['img_dirs'])
image = tf.image.decode_image(img_file, channels=3)
image = tf.image.pad_to_bounding_box(image, 0, 0, imh, imw)
image.set_shape((imh, imw, 3))
image = tf.cast(image, tf.float32) / 255.0 - 0.5
data_dict['image'] = image
if 'mask_dirs' in flow_dict:
mask_file = tf.read_file(flow_dict['mask_dirs'])
mask = tf.image.decode_image(mask_file, channels=3)
mask = tf.image.pad_to_bounding_box(mask, 0, 0, imh, imw)
mask.set_shape((imh, imw, 3))
mask = mask[:, :, 0]
mask = tf.cast(mask, tf.float32)
else:
mask = tf.ones((imh, imw), dtype=tf.float32)
data_dict['mask'] = mask
# calculate crop size
if self.objtype in (0, 1):
if self.objtype == 0:
keypoints = body3d
valid = flow_dict['body_valid']
elif self.objtype == 1:
keypoints = hand3d
valid = tf.cond(cond_left, lambda: flow_dict['left_hand_valid'], lambda: flow_dict['right_hand_valid'])
data_dict['hand_valid'] = valid
crop_center3d, scale3d, crop_center2d, scale2d = self.calc_crop_scale(keypoints, flow_dict['K'], flow_dict['distCoef'], valid)
data_dict['crop_center2d'], data_dict['scale2d'] = crop_center2d, scale2d
data_dict['crop_center3d'], data_dict['scale3d'] = crop_center3d, scale3d
# do cropping
if self.objtype == 1:
body2d = hand2d
body3d = hand3d
if self.rotate_augmentation:
print('using rotation augmentation')
rotate_angle = tf.random_uniform([], minval=-np.pi * 40 / 180, maxval=np.pi * 40 / 180)
R2 = tf.reshape(tf.stack([tf.cos(rotate_angle), -tf.sin(rotate_angle), tf.sin(rotate_angle), tf.cos(rotate_angle)]), [2, 2])
R3 = tf.reshape(tf.stack([tf.cos(rotate_angle), -tf.sin(rotate_angle), 0, tf.sin(rotate_angle), tf.cos(rotate_angle), 0, 0, 0, 1]), [3, 3])
body2d = tf.matmul((body2d - crop_center2d), R2) + crop_center2d
body3d = tf.matmul((body3d - crop_center3d), R3) + crop_center3d
data_dict['keypoint_xyz_origin'] = body3d # note that the projection of 3D might not be aligned with 2D any more after rotation
data_dict['keypoint_uv_origin'] = body2d
body2d_local = self.update_keypoint2d(body2d, crop_center2d, scale2d)
data_dict['keypoint_uv_local'] = body2d_local
if read_image:
image_crop = self.crop_image(image, crop_center2d, scale2d)
data_dict['image_crop'] = image_crop
mask_crop = self.crop_image(tf.stack([mask] * 3, axis=2), crop_center2d, scale2d)
data_dict['mask_crop'] = mask_crop[:, :, 0]
if self.rotate_augmentation:
data_dict['image_crop'] = tf.contrib.image.rotate(data_dict['image_crop'], rotate_angle)
data_dict['mask_crop'] = tf.contrib.image.rotate(data_dict['mask_crop'], rotate_angle)
if self.blur_augmentation:
print('using blur augmentation')
rescale_factor = tf.random_uniform([], minval=0.1, maxval=1.0)
rescale = tf.cast(rescale_factor * self.crop_size, tf.int32)
resized_image = tf.image.resize_images(data_dict['image_crop'], [rescale, rescale])
data_dict['image_crop'] = tf.image.resize_images(resized_image, [self.crop_size, self.crop_size])
# create 2D gaussian map
scoremap2d = self.create_multiple_gaussian_map(body2d_local[:, ::-1], (self.crop_size, self.crop_size), self.sigma, valid_vec=valid, extra=True) # coord_hw, imsize_hw
data_dict['scoremap2d'] = scoremap2d
if withPAF:
from utils.PAF import createPAF
data_dict['PAF'] = createPAF(body2d_local, body3d, self.objtype, (self.crop_size, self.crop_size), PAF_normalize3d, valid_vec=valid)
data_dict['PAF_type'] = tf.ones([], dtype=bool) # 0 for 2D PAF, 1 for 3D PAF
# create 3D gaussian_map
body3d_local = self.update_keypoint3d(body3d, crop_center3d, scale3d)
data_dict['keypoint_xyz_local'] = body3d_local
# scoremap3d = self.create_multiple_gaussian_map_3d(body3d_local, self.grid_size, self.sigma3d, valid_vec=valid, extra=True)
# data_dict['scoremap3d'] = scoremap3d
if self.objtype == 1: # this is hand, flip the image if it is right hand
data_dict['image_crop'] = tf.cond(cond_left, lambda: data_dict['image_crop'], lambda: data_dict['image_crop'][:, ::-1, :])
data_dict['mask_crop'] = tf.cond(cond_left, lambda: data_dict['mask_crop'], lambda: data_dict['mask_crop'][:, ::-1])
data_dict['scoremap2d'] = tf.cond(cond_left, lambda: data_dict['scoremap2d'], lambda: data_dict['scoremap2d'][:, ::-1, :])
data_dict['keypoint_uv_local'] = tf.cond(cond_left, lambda: data_dict['keypoint_uv_local'],
lambda: tf.constant([self.crop_size, 0], tf.float32) + tf.constant([-1, 1], tf.float32) * data_dict['keypoint_uv_local'])
if withPAF:
data_dict['PAF'] = tf.cond(cond_left, lambda: data_dict['PAF'],
lambda: (data_dict['PAF'][:, ::-1, :]) * tf.constant([-1, 1, 1] * (data_dict['PAF'].get_shape().as_list()[2] // 3), dtype=tf.float32))
elif self.objtype == 2:
bcrop_center3d, bscale3d, bcrop_center2d, bscale2d = self.calc_crop_scale(body3d, flow_dict['K'], flow_dict['distCoef'], flow_dict['body_valid'])
lcrop_center3d, lscale3d, lcrop_center2d, lscale2d = self.calc_crop_scale(lhand3d, flow_dict['K'], flow_dict['distCoef'], flow_dict['left_hand_valid'])
rcrop_center3d, rscale3d, rcrop_center2d, rscale2d = self.calc_crop_scale(rhand3d, flow_dict['K'], flow_dict['distCoef'], flow_dict['right_hand_valid'])
body3d_local = self.update_keypoint3d(body3d, bcrop_center3d, bscale3d)
lhand3d_local = self.update_keypoint3d(lhand3d, lcrop_center3d, lscale3d)
rhand3d_local = self.update_keypoint3d(rhand3d, rcrop_center3d, rscale3d)
bscoremap3d = self.create_multiple_gaussian_map_3d(body3d_local, self.grid_size, self.sigma3d,
valid_vec=flow_dict['body_valid'], extra=True) # coord_hw, imsize_hw
lscoremap3d = self.create_multiple_gaussian_map_3d(lhand3d_local, self.grid_size, self.sigma3d,
valid_vec=flow_dict['left_hand_valid'], extra=True) # coord_hw, imsize_hw
rscoremap3d = self.create_multiple_gaussian_map_3d(rhand3d_local, self.grid_size, self.sigma3d,
valid_vec=flow_dict['right_hand_valid'], extra=True) # coord_hw, imsize_hw
data_dict['bscoremap3d'] = bscoremap3d
data_dict['lscoremap3d'] = lscoremap3d
data_dict['rscoremap3d'] = rscoremap3d
data_dict['body_xyz_local'] = body3d_local
data_dict['lhand_xyz_local'] = lhand3d_local
data_dict['rhand_xyz_local'] = rhand3d_local
# 2D keypoints and cropped images
if bbox2d == 1:
# crop the 2D bounding box from openpose data
body2d = flow_dict['openpose_body']
lhand2d = flow_dict['openpose_lhand']
rhand2d = flow_dict['openpose_rhand']
bvalid = tf.logical_and(tf.not_equal(body2d[:, 0], 0.0), tf.not_equal(body2d[:, 1], 0.0))
lvalid = tf.logical_and(tf.not_equal(lhand2d[:, 0], 0.0), tf.not_equal(lhand2d[:, 1], 0.0))
rvalid = tf.logical_and(tf.not_equal(rhand2d[:, 0], 0.0), tf.not_equal(rhand2d[:, 1], 0.0))
data_dict['body_valid'] = bvalid
data_dict['left_hand_valid'] = lvalid
data_dict['right_hand_valid'] = rvalid
if 'openpose_foot' in flow_dict:
data_dict['openpose_foot'] = flow_dict['openpose_foot']
bcrop_center2d, bscale2d = self.calc_crop_scale2d(body2d, bvalid)
lcrop_center2d, lscale2d = self.calc_crop_scale2d(lhand2d, lvalid)
rcrop_center2d, rscale2d = self.calc_crop_scale2d(rhand2d, rvalid)
body2d_local = self.update_keypoint2d(body2d, bcrop_center2d, bscale2d)
lhand2d_local = self.update_keypoint2d(lhand2d, lcrop_center2d, lscale2d)
rhand2d_local = self.update_keypoint2d(rhand2d, rcrop_center2d, rscale2d)
data_dict['body_uv_local'] = body2d_local
data_dict['lhand_uv_local'] = lhand2d_local
data_dict['rhand_uv_local'] = rhand2d_local
data_dict['bcrop_center2d'] = bcrop_center2d
data_dict['lcrop_center2d'] = lcrop_center2d
data_dict['rcrop_center2d'] = rcrop_center2d
data_dict['bscale2d'] = bscale2d
data_dict['lscale2d'] = lscale2d
data_dict['rscale2d'] = rscale2d
if read_image:
bimage_crop = self.crop_image(image, bcrop_center2d, bscale2d)
limage_crop = self.crop_image(image, lcrop_center2d, lscale2d)
rimage_crop = self.crop_image(image, rcrop_center2d, rscale2d)
data_dict['bimage_crop'] = bimage_crop
data_dict['limage_crop'] = limage_crop
data_dict['rimage_crop'] = rimage_crop
bscoremap2d = self.create_multiple_gaussian_map(body2d_local[:, ::-1], (self.crop_size, self.crop_size), self.sigma,
valid_vec=flow_dict['body_valid'], extra=True) # coord_hw, imsize_hw
lscoremap2d = self.create_multiple_gaussian_map(lhand2d_local[:, ::-1], (self.crop_size, self.crop_size), self.sigma,
valid_vec=flow_dict['left_hand_valid'], extra=True) # coord_hw, imsize_hw
rscoremap2d = self.create_multiple_gaussian_map(rhand2d_local[:, ::-1], (self.crop_size, self.crop_size), self.sigma,
valid_vec=flow_dict['right_hand_valid'], extra=True) # coord_hw, imsize_hw
data_dict['bscoremap2d'] = bscoremap2d
data_dict['lscoremap2d'] = lscoremap2d
data_dict['rscoremap2d'] = rscoremap2d
# for openpose data
for key, val in flow_dict.items():
if 'openpose' not in key:
continue
data_dict[key] = val
names, tensors = zip(*data_dict.items())
if self.shuffle:
tensors = tf.train.shuffle_batch_join([tensors],
batch_size=self.batch_size,
capacity=100,
min_after_dequeue=20,
enqueue_many=False)
else:
tensors = tf.train.batch_join([tensors],
batch_size=self.batch_size,
capacity=20,
enqueue_many=False)
return dict(zip(names, tensors))