in visualization/POF/data/TempConstReader.py [0:0]
def get(self, withPAF=True, read_image=True, imw=1920, imh=1080):
# input to this data reader should have two consecutive frames
# 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['1_img_dir'] = flow_dict['1_img_dirs']
data_dict['2_img_dir'] = flow_dict['2_img_dirs']
data_dict['1_K'] = flow_dict['1_K']
data_dict['1_K'] = flow_dict['2_K']
# rotate and project to camera frame
if self.objtype == 0:
body2d_1, body3d_1 = self.project_tf(flow_dict['1_body'], flow_dict['1_K'], flow_dict['1_R'], flow_dict['1_t'], flow_dict['1_distCoef'])
body2d_2, body3d_2 = self.project_tf(flow_dict['2_body'], flow_dict['2_K'], flow_dict['2_R'], flow_dict['2_t'], flow_dict['2_distCoef'])
body3d_1 = tf.cast(body3d_1, tf.float32)
body3d_2 = tf.cast(body3d_2, tf.float32)
body2d_1 = tf.cast(body2d_1, tf.float32)
body2d_2 = tf.cast(body2d_2, tf.float32)
data_dict['1_keypoint_xyz_origin'] = body3d_1
data_dict['2_keypoint_xyz_origin'] = body3d_2
data_dict['1_keypoint_uv_origin'] = body2d_1
data_dict['2_keypoint_uv_origin'] = body2d_2
data_dict['1_body_valid'] = flow_dict['1_body_valid']
data_dict['2_body_valid'] = flow_dict['2_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_1 = tf.cond(cond_left, lambda: flow_dict['1_left_hand'], lambda: flow_dict['1_right_hand']) # in world coordinate
hand3d_2 = tf.cond(cond_left, lambda: flow_dict['2_left_hand'], lambda: flow_dict['2_right_hand']) # in world coordinate
hand2d_1, hand3d_1 = self.project_tf(hand3d_1, flow_dict['1_K'], flow_dict['1_R'], flow_dict['1_t'], flow_dict['1_distCoef']) # in camera coordinate
hand2d_2, hand3d_2 = self.project_tf(hand3d_2, flow_dict['2_K'], flow_dict['2_R'], flow_dict['2_t'], flow_dict['2_distCoef']) # in camera coordinate
hand3d_1 = tf.cast(hand3d_1, tf.float32)
hand3d_2 = tf.cast(hand3d_2, tf.float32)
hand2d_1 = tf.cast(hand2d_1, tf.float32)
hand2d_2 = tf.cast(hand2d_2, tf.float32)
data_dict['1_keypoint_xyz_origin'] = hand3d_1
data_dict['2_keypoint_xyz_origin'] = hand3d_2
data_dict['1_keypoint_uv_origin'] = hand2d_1
data_dict['2_keypoint_uv_origin'] = hand2d_2
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']
# read image
if read_image:
img_file_1 = tf.read_file(flow_dict['1_img_dirs'])
img_file_2 = tf.read_file(flow_dict['2_img_dirs'])
image_1 = tf.image.decode_image(img_file_1, channels=3)
image_2 = tf.image.decode_image(img_file_2, channels=3)
image_1 = tf.image.pad_to_bounding_box(image_1, 0, 0, imh, imw)
image_2 = tf.image.pad_to_bounding_box(image_2, 0, 0, imh, imw)
image_1.set_shape((imh, imw, 3))
image_2.set_shape((imh, imw, 3))
image_1 = tf.cast(image_1, tf.float32) / 255.0 - 0.5
image_2 = tf.cast(image_2, tf.float32) / 255.0 - 0.5
data_dict['1_image'] = image_1
data_dict['2_image'] = image_2
if 'mask_dirs_1' in flow_dict:
assert 'mask_dirs_2' in flow_dict
mask_file_1 = tf.read_file(flow_dict['1_mask_dirs'])
mask_file_2 = tf.read_file(flow_dict['2_mask_dirs'])
mask_1 = tf.image.decode_image(mask_file_1, channels=3)
mask_2 = tf.image.decode_image(mask_file_2, channels=3)
mask_1 = tf.image.pad_to_bounding_box(mask_1, 0, 0, imh, imw)
mask_2 = tf.image.pad_to_bounding_box(mask_2, 0, 0, imh, imw)
mask_1.set_shape((imh, imw, 3))
mask_2.set_shape((imh, imw, 3))
mask_1 = mask_1[:, :, 0]
mask_2 = mask_2[:, :, 0]
mask_1 = tf.cast(mask_1, tf.float32)
mask_2 = tf.cast(mask_2, tf.float32)
else:
mask_1 = tf.ones((imh, imw), dtype=tf.float32)
mask_2 = tf.ones((imh, imw), dtype=tf.float32)
data_dict['1_mask'] = mask_1
data_dict['2_mask'] = mask_2
# calculate crop size
if self.objtype in (0, 1):
if self.objtype == 0:
keypoints_1 = body3d_1
keypoints_2 = body3d_2
valid_1 = flow_dict['1_body_valid']
valid_2 = flow_dict['2_body_valid']
elif self.objtype == 1:
keypoints_1 = hand3d_1
keypoints_2 = hand3d_2
valid_1 = tf.cond(cond_left, lambda: flow_dict['left_hand_valid'], lambda: flow_dict['right_hand_valid'])
valid_2 = tf.cond(cond_left, lambda: flow_dict['left_hand_valid'], lambda: flow_dict['right_hand_valid'])
data_dict['1_hand_valid'] = valid_1
data_dict['2_hand_valid'] = valid_2
crop_center3d_1, scale3d_1, crop_center2d_1, scale2d_1, crop_center3d_2, scale3d_2, crop_center2d_2, scale2d_2 = \
self.calc_crop_scale_temp_const(keypoints_1, flow_dict['1_K'], flow_dict['1_distCoef'], valid_1, keypoints_2, flow_dict['2_K'], flow_dict['2_distCoef'], valid_2)
data_dict['1_crop_center2d'], data_dict['1_scale2d'] = crop_center2d_1, scale2d_1
data_dict['2_crop_center2d'], data_dict['2_scale2d'] = crop_center2d_2, scale2d_2
data_dict['1_crop_center3d'], data_dict['1_scale3d'] = crop_center3d_1, scale3d_1
data_dict['2_crop_center3d'], data_dict['2_scale3d'] = crop_center3d_2, scale3d_2
# do cropping
if self.objtype == 1:
body2d_1 = hand2d_1
body2d_2 = hand2d_2
body3d_1 = hand3d_1
body3d_2 = hand3d_2
if self.rotate_augmentation:
print('using rotation augmentation')
rotate_angle_1 = tf.random_uniform([], minval=-np.pi * 40 / 180, maxval=np.pi * 40 / 180)
else:
rotate_angle_1 = 0.0
rotate_angle_2 = tf.random_uniform([], minval=-np.pi * 5 / 180, maxval=np.pi * 5 / 180) + rotate_angle_1
R2_1 = tf.reshape(tf.stack([tf.cos(rotate_angle_1), -tf.sin(rotate_angle_1), tf.sin(rotate_angle_1), tf.cos(rotate_angle_1)]), [2, 2])
R2_2 = tf.reshape(tf.stack([tf.cos(rotate_angle_2), -tf.sin(rotate_angle_2), tf.sin(rotate_angle_2), tf.cos(rotate_angle_2)]), [2, 2])
R3_1 = tf.reshape(tf.stack([tf.cos(rotate_angle_1), -tf.sin(rotate_angle_1), 0, tf.sin(rotate_angle_1), tf.cos(rotate_angle_1), 0, 0, 0, 1]), [3, 3])
R3_2 = tf.reshape(tf.stack([tf.cos(rotate_angle_2), -tf.sin(rotate_angle_2), 0, tf.sin(rotate_angle_2), tf.cos(rotate_angle_2), 0, 0, 0, 1]), [3, 3])
body2d_1 = tf.matmul((body2d_1 - crop_center2d_1), R2_1) + crop_center2d_1
body2d_2 = tf.matmul((body2d_2 - crop_center2d_2), R2_2) + crop_center2d_2
body3d_1 = tf.matmul((body3d_1 - crop_center3d_1), R3_1) + crop_center3d_1
body3d_2 = tf.matmul((body3d_2 - crop_center3d_2), R3_2) + crop_center3d_2
data_dict['1_keypoint_xyz_origin'] = body3d_1 # note that the projection of 3D might not be aligned with 2D any more after rotation
data_dict['2_keypoint_xyz_origin'] = body3d_2 # note that the projection of 3D might not be aligned with 2D any more after rotation
data_dict['1_keypoint_uv_origin'] = body2d_1
data_dict['2_keypoint_uv_origin'] = body2d_2
body2d_local_1 = self.update_keypoint2d(body2d_1, crop_center2d_1, scale2d_1)
body2d_local_2 = self.update_keypoint2d(body2d_2, crop_center2d_2, scale2d_2)
data_dict['1_keypoint_uv_local'] = body2d_local_1
data_dict['2_keypoint_uv_local'] = body2d_local_2
if read_image:
image_crop_1 = self.crop_image(image_1, crop_center2d_1, scale2d_1)
image_crop_2 = self.crop_image(image_2, crop_center2d_2, scale2d_2)
data_dict['1_image_crop'] = image_crop_1
data_dict['2_image_crop'] = image_crop_2
mask_crop_1 = self.crop_image(tf.stack([mask_1] * 3, axis=2), crop_center2d_1, scale2d_1)
mask_crop_2 = self.crop_image(tf.stack([mask_2] * 3, axis=2), crop_center2d_2, scale2d_2)
data_dict['1_mask_crop'] = mask_crop_1[:, :, 0]
data_dict['2_mask_crop'] = mask_crop_2[:, :, 0]
data_dict['1_image_crop'] = tf.contrib.image.rotate(data_dict['1_image_crop'], rotate_angle_1)
data_dict['2_image_crop'] = tf.contrib.image.rotate(data_dict['2_image_crop'], rotate_angle_2)
data_dict['1_mask_crop'] = tf.contrib.image.rotate(data_dict['1_mask_crop'], rotate_angle_1)
data_dict['2_mask_crop'] = tf.contrib.image.rotate(data_dict['2_mask_crop'], rotate_angle_2)
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_1 = tf.image.resize_images(data_dict['1_image_crop'], [rescale, rescale])
resized_image_2 = tf.image.resize_images(data_dict['2_image_crop'], [rescale, rescale])
data_dict['1_image_crop'] = tf.image.resize_images(resized_image_1, [self.crop_size, self.crop_size])
data_dict['2_image_crop'] = tf.image.resize_images(resized_image_2, [self.crop_size, self.crop_size])
# create 2D gaussian map
scoremap2d_1 = self.create_multiple_gaussian_map(body2d_local_1[:, ::-1], (self.crop_size, self.crop_size), self.sigma, valid_vec=valid_1, extra=True) # coord_hw, imsize_hw
scoremap2d_2 = self.create_multiple_gaussian_map(body2d_local_2[:, ::-1], (self.crop_size, self.crop_size), self.sigma, valid_vec=valid_2, extra=True) # coord_hw, imsize_hw
data_dict['1_scoremap2d'] = scoremap2d_1
data_dict['2_scoremap2d'] = scoremap2d_2
if withPAF:
from utils.PAF import createPAF
data_dict['1_PAF'] = createPAF(body2d_local_1, body3d_1, self.objtype, (self.crop_size, self.crop_size), True, valid_vec=valid_1)
data_dict['2_PAF'] = createPAF(body2d_local_2, body3d_2, self.objtype, (self.crop_size, self.crop_size), True, valid_vec=valid_2)
data_dict['1_PAF_type'] = tf.ones([], dtype=bool) # 0 for 2D PAF, 1 for 3D PAF
data_dict['2_PAF_type'] = tf.ones([], dtype=bool) # 0 for 2D PAF, 1 for 3D PAF
# create 3D gaussian_map
body3d_local_1 = self.update_keypoint3d(body3d_1, crop_center3d_1, scale3d_1)
body3d_local_2 = self.update_keypoint3d(body3d_2, crop_center3d_2, scale3d_2)
data_dict['1_keypoint_xyz_local'] = body3d_local_1
data_dict['2_keypoint_xyz_local'] = body3d_local_2
# scoremap3d = self.create_multiple_gaussian_map_3d(body3d_local, self.grid_size, self.sigma3d, valid_vec=valid, extra=True)
# data_dict['1_scoremap3d'] = scoremap3d
if self.objtype == 1: # this is hand, flip the image if it is right hand
data_dict['1_image_crop'] = tf.cond(cond_left, lambda: data_dict['1_image_crop'], lambda: data_dict['1_image_crop'][:, ::-1, :])
data_dict['2_image_crop'] = tf.cond(cond_left, lambda: data_dict['2_image_crop'], lambda: data_dict['2_image_crop'][:, ::-1, :])
data_dict['1_mask_crop'] = tf.cond(cond_left, lambda: data_dict['1_mask_crop'], lambda: data_dict['1_mask_crop'][:, ::-1])
data_dict['2_mask_crop'] = tf.cond(cond_left, lambda: data_dict['2_mask_crop'], lambda: data_dict['2_mask_crop'][:, ::-1])
data_dict['1_scoremap2d'] = tf.cond(cond_left, lambda: data_dict['1_scoremap2d'], lambda: data_dict['1_scoremap2d'][:, ::-1, :])
data_dict['2_scoremap2d'] = tf.cond(cond_left, lambda: data_dict['2_scoremap2d'], lambda: data_dict['2_scoremap2d'][:, ::-1, :])
data_dict['1_keypoint_uv_local'] = tf.cond(cond_left, lambda: data_dict['1_keypoint_uv_local'],
lambda: tf.constant([self.crop_size, 0], tf.float32) + tf.constant([-1, 1], tf.float32) * data_dict['1_keypoint_uv_local'])
data_dict['2_keypoint_uv_local'] = tf.cond(cond_left, lambda: data_dict['2_keypoint_uv_local'],
lambda: tf.constant([self.crop_size, 0], tf.float32) + tf.constant([-1, 1], tf.float32) * data_dict['2_keypoint_uv_local'])
if withPAF:
data_dict['1_PAF'] = tf.cond(cond_left, lambda: data_dict['1_PAF'],
lambda: (data_dict['1_PAF'][:, ::-1, :]) * tf.constant([-1, 1, 1] * (data_dict['1_PAF'].get_shape().as_list()[2] // 3), dtype=tf.float32))
data_dict['2_PAF'] = tf.cond(cond_left, lambda: data_dict['2_PAF'],
lambda: (data_dict['2_PAF'][:, ::-1, :]) * tf.constant([-1, 1, 1] * (data_dict['2_PAF'].get_shape().as_list()[2] // 3), dtype=tf.float32))
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))