def loadBatchSurreal_fromString(file_string, image_size=128, num_frames=2, \
keypoints_num=24, bases_num=10, chamfer_scale=0.5):
filename, t = file_string.split("#")
output = dict()
output[0] = read_syn_to_bin2(filename, int(t))
#print "start!!@@@@@"
scipy.io.savemat('tf_test1.mat', {'output': output[0]})
output[1] = read_syn_to_bin2(filename, int(t) + 1)
#output[0] = read_syn_to_bin(filename, int(t))
#output[1] = read_syn_to_bin(filename, int(t) + 1)
data_pose = np.zeros((num_frames, keypoints_num * 3))
data_T = np.zeros((num_frames, 3))
data_R = np.zeros((num_frames, 6))
data_beta = np.zeros((num_frames, bases_num))
data_J = np.zeros((num_frames, keypoints_num, 3))
data_J_2d = np.zeros((num_frames, keypoints_num, 2))
data_image = np.zeros((num_frames, image_size, image_size, 3))
data_seg = np.zeros((num_frames, image_size, image_size))
small_image_size = int(chamfer_scale * image_size)
data_chamfer = np.zeros((num_frames, small_image_size, small_image_size))
data_f = np.zeros((num_frames, 2))
data_c = np.zeros((num_frames, 2))
data_resize_scale = np.zeros((num_frames))
data_gender = np.zeros(())
# Cropping
old_2d_center = output[0]['c'] #np.array([(320 - 1)/2.0, (240-1)/2.0])
# Use keypoint 0 in frame1 as center
J_2d = output[0]['J_2d']
new_2d_center = np.round(J_2d[0, :]) + 0.5 * np.ones((2))
s = 1.2 #1.3 + 0.1 * np.random.rand()
crop_size = np.round(
s * np.max(np.abs(J_2d - np.reshape(new_2d_center, [1, 1, -1]))))
new_image_size = int(2 * crop_size)
x_min = int(math.ceil(new_2d_center[0] - crop_size))
x_max = int(math.floor(new_2d_center[0] + crop_size))
y_min = int(math.ceil(new_2d_center[1] - crop_size))
y_max = int(math.floor(new_2d_center[1] + crop_size))
resize_scale = float(image_size) / (crop_size * 2.0)
data_resize_scale[:] = resize_scale
new_origin = np.array([x_min, y_min])
data_c[:, :] = np.reshape(old_2d_center - new_origin, [-1, 2])
for frame_id in range(num_frames):
data_pose[frame_id, :] = output[frame_id]['pose']
data_gender = int(output[frame_id]['gender'])
data_R[frame_id, :3] = np.sin(output[frame_id]['R'])
data_R[frame_id, 3:6] = np.cos(output[frame_id]['R'])
#data_beta[frame_id, :] = output[frame_id]['beta']
data_f[frame_id, :] = output[frame_id]['f']
#data_c[frame_id, :] = output[frame_id]['c']
data_T[frame_id, :] = output[frame_id]['T']
data_J[frame_id, :, :] = output[frame_id]['J']
data_J_2d[frame_id, :, :] = resize_scale * (
output[frame_id]['J_2d'] - np.reshape(new_origin, [1, -1]))
# crop image
image = output[frame_id]['image']
#print image
h, w, _ = image.shape
img_x_min = max(x_min, 0)
img_x_max = min(x_max, w - 1)
img_y_min = max(y_min, 0)
img_y_max = min(y_max, h - 1)
crop_image = np.zeros((new_image_size, new_image_size, 3),
dtype=np.float32)
crop_image[max(0, -y_min):max(0, -y_min) + img_y_max - img_y_min + 1, \
max(0, -x_min):max(0, -x_min) + img_x_max - img_x_min + 1, :] \
= image[img_y_min:img_y_max + 1, img_x_min:img_x_max +1, :]
data_image[frame_id, :, :, :] = scipy.misc.imresize(
crop_image, [image_size, image_size])
seg_float = output[frame_id]['seg'].astype(np.float32)
crop_seg = np.zeros((new_image_size, new_image_size, 3),
dtype=np.float32)
crop_seg[max(0, -y_min):max(0, -y_min) + img_y_max - img_y_min + 1, \
max(0, -x_min):max(0, -x_min) + img_x_max - img_x_min + 1] \
= np.expand_dims(seg_float[img_y_min:img_y_max + 1, \
img_x_min:img_x_max +1], 2)
seg = scipy.misc.imresize(crop_seg, [image_size, image_size])
seg[seg < 0.5] = 0
seg[seg >= 0.5] = 1
#print np.max(output['seg'][sample_id, :, :])
data_seg[frame_id, :, :] = seg[:, :, 0]
data_chamfer[frame_id, :, :], _, _ = get_chamfer(
seg[:, :, 0], chamfer_scale)
# return data_pose, data_T, data_R, data_J, data_J_2d,\
# data_f, data_c, data_gender
#print "resize",data_resize_scale*3.0
return data_pose, data_T, data_R, data_beta, data_J, data_J_2d, data_image/255.0,\
data_seg, data_f/3.0, data_chamfer, data_c/3.0, data_gender, data_resize_scale*3.0
def next(self):
if self.batch_id >= self.num_batches:
self.reset()
images_perm = self.perm[int(self.batch_id *
self.batch_size):int((self.batch_id + 1) *
self.batch_size)]
#batch_images = self.images[images_perm,:,:,:]
batch_pose = np.zeros(
(self.batch_size, self.num_frames, self.keypoints_num, 3),
dtype=np.float32)
batch_T = np.zeros((self.batch_size, self.num_frames, 3),
dtype=np.float32)
batch_R = np.zeros((self.batch_size, self.num_frames, 6),
dtype=np.float32)
batch_gender = np.zeros((self.batch_size), dtype=np.int32)
batch_beta = np.zeros(
(self.batch_size, self.num_frames, self.bases_num),
dtype=np.float32)
batch_J = np.zeros(
(self.batch_size, self.num_frames, self.keypoints_num, 3),
dtype=np.float32)
batch_J_2d = np.zeros(
(self.batch_size, self.num_frames, self.keypoints_num, 2),
dtype=np.float32)
batch_image = np.zeros(
(self.batch_size, self.num_frames, self.h, self.w, 3),
dtype=np.float32)
batch_seg = np.zeros(
(self.batch_size, self.num_frames, self.h, self.w),
dtype=np.float32)
batch_chamfer = np.zeros(
(self.batch_size, self.num_frames, self.small_h, self.small_w),
dtype=np.float32)
batch_f = np.zeros((self.batch_size, self.num_frames, 2),
dtype=np.float32)
batch_c = np.zeros((self.batch_size, self.num_frames, 2),
dtype=np.float32)
batch_resize_scale = np.zeros((self.batch_size, self.num_frames),
dtype=np.float32)
old_2d_center = np.array([(320 - 1) / 2.0, (240 - 1) / 2.0])
for i in range(self.batch_size):
id_ = images_perm[i]
#batch_labels_2d[i,:,0] = self.labels_2d[id_, 0:self.keypoints_num]
#batch_labels_2d[i,:,1] = self.labels_2d[id_, self.keypoints_num:self.keypoints_num*2]
batch_pose[i, :, :, :] = self.data_pose[id_, :, :, :]
batch_gender[i] = self.data_gender[id_]
batch_R[i, :, :3] = np.sin(self.data_R[id_, :, :])
batch_R[i, :, 3:6] = np.cos(self.data_R[id_, :, :])
batch_beta[i, :, :] = self.data_beta[id_, :, :]
batch_f[i, :, :] = self.data_f[id_, :, :]
batch_T[i, :, :] = self.data_T[id_, :, :]
batch_J[i, :, :, :] = self.data_J[id_, :, :, :]
J_2d = self.data_J_2d[id_, :, :, :]
new_2d_center = np.round(J_2d[0, 0, :] + 10 * (np.random.uniform(
(2)) - 1)) + 0.5 * np.ones((2))
crop_size = np.round(
1.2 *
np.max(np.abs(J_2d - np.reshape(new_2d_center, [1, 1, -1]))))
new_image_size = int(2 * crop_size)
x_min = int(math.ceil(new_2d_center[0] - crop_size))
x_max = int(math.floor(new_2d_center[0] + crop_size))
y_min = int(math.ceil(new_2d_center[1] - crop_size))
y_max = int(math.floor(new_2d_center[1] + crop_size))
resize_scale = self.h / (crop_size * 2.0)
batch_resize_scale[i, :] = resize_scale
# frame_id * 2
new_origin = np.array([x_min, y_min])
batch_c[i, :, :] = np.reshape(old_2d_center - new_origin, [-1, 2])
batch_J_2d[i, :, :, :] = resize_scale * (
self.data_J_2d[id_, :, :, :] -
np.reshape(new_origin, [1, 1, -1]))
image = self.data_image[id_, :, :, :, :].astype(np.float32)
seg_float = self.data_seg[id_, :, :, :].astype(np.float32)
img_x_min = max(x_min, 0)
img_x_max = min(x_max, 359)
img_y_min = max(y_min, 0)
img_y_max = min(y_max, 239)
#print new_2d_center
#print crop_size
print(img_x_min, img_x_max, img_y_min, img_y_max)
print(x_min, x_max, y_min, y_max)
for frame_id in range(self.num_frames):
crop_image = np.zeros((new_image_size, new_image_size, 3),
dtype=np.float32)
crop_image[max(0, -y_min):max(0, -y_min) + img_y_max - img_y_min + 1, \
max(0, -x_min):max(0, -x_min) + img_x_max - img_x_min + 1, :] \
= image[frame_id, img_y_min:img_y_max + 1, img_x_min:img_x_max +1, :]
batch_image[i, frame_id, :, :, :] = scipy.misc.imresize(
crop_image, [self.h, self.w])
crop_seg = np.zeros((new_image_size, new_image_size, 3),
dtype=np.float32)
crop_seg[max(0, -y_min):max(0, -y_min) + img_y_max - img_y_min + 1, \
max(0, -x_min):max(0, -x_min) + img_x_max - img_x_min + 1] \
= np.expand_dims(seg_float[frame_id, img_y_min:img_y_max + 1, \
img_x_min:img_x_max +1], 2)
seg = scipy.misc.imresize(crop_seg, [self.h, self.w])
seg[seg < 0.5] = 0
seg[seg >= 0.5] = 1
batch_seg[i, frame_id, :, :] = seg[:, :, 0]
# calculate chamfer dist
#for img_id in range(self.num_frames):
batch_chamfer[i, frame_id, :, :], _, _ = get_chamfer(
seg[:, :, 0], self.chamfer_scale)
self.batch_id += 1
return batch_pose, batch_T, batch_R, batch_beta, batch_J, batch_J_2d, batch_image/255.0,\
batch_seg, batch_f, batch_chamfer, batch_c, batch_gender, batch_resize_scale
def loadBatchSurreal_fromString(file_string,
image_size=128,
num_frames=2,
keypoints_num=24,
bases_num=10,
chamfer_scale=0.5,
is_gait=False):
filename, t = file_string.split("#")
output = dict()
output[0] = read_syn_to_bin(filename, int(t))
output[1] = read_syn_to_bin(filename, int(t) + 1)
data_pose = np.zeros((num_frames, keypoints_num * 3))
data_T = np.zeros((num_frames, 3))
data_R = np.zeros((num_frames, 6))
data_beta = np.zeros((num_frames, bases_num))
data_J = np.zeros((num_frames, keypoints_num, 3))
data_J_2d = np.zeros((num_frames, keypoints_num, 2))
data_J_reconstruct_2d = np.zeros((num_frames, keypoints_num, 2))
data_image = np.zeros((num_frames, image_size, image_size, 3))
data_seg = np.zeros((num_frames, image_size, image_size))
small_image_size = int(chamfer_scale * image_size)
data_chamfer = np.zeros((num_frames, small_image_size, small_image_size))
data_f = np.zeros((num_frames, 2))
data_c = np.zeros((num_frames, 2))
data_resize_scale = np.zeros((num_frames))
data_gender = np.zeros(())
# Cropping
w, h = (320, 180)
old_2d_center = np.array([(w - 1) / 2.0, (h - 1) / 2.0])
# Use keypoint 0 in frame1 as center
J_2d = output[0]['J_2d']
new_2d_center = np.round(J_2d[0, :] + 10 *
(np.random.uniform((2)) - 1)) + 0.5 * np.ones((2))
s = 1.2 #+ 0.1 * np.random.rand()
crop_size = np.round(
s * np.max(np.abs(J_2d - np.reshape(new_2d_center, [1, 1, -1]))))
new_image_size = int(2 * crop_size)
x_min = int(math.ceil(new_2d_center[0] - crop_size))
x_max = int(math.floor(new_2d_center[0] + crop_size))
y_min = int(math.ceil(new_2d_center[1] - crop_size))
y_max = int(math.floor(new_2d_center[1] + crop_size))
resize_scale = float(image_size) / (crop_size * 2.0)
data_resize_scale[:] = resize_scale
new_origin = np.array([x_min, y_min])
data_c[:, :] = np.reshape(old_2d_center - new_origin, [-1, 2])
for frame_id in range(num_frames):
data_pose[frame_id, :] = output[frame_id]['pose']
data_gender = int(output[frame_id]['gender'])
data_R[frame_id, :3] = np.sin(output[frame_id]['R'])
data_R[frame_id, 3:6] = np.cos(output[frame_id]['R'])
data_beta[frame_id, :] = output[frame_id]['beta']
data_f[frame_id, :] = output[frame_id]['f']
data_T[frame_id, :] = output[frame_id]['T']
data_J[frame_id, :, :] = output[frame_id]['J']
data_J_2d[frame_id, :, :] = resize_scale * (
output[frame_id]['J_2d'] - np.reshape(new_origin, [1, -1]))
data_J_reconstruct_2d[frame_id, :, :] = resize_scale * (
output[frame_id]['J_reconstruct_2d'] -
np.reshape(new_origin, [1, -1]))
# crop image
image = output[frame_id]['image']
h, w, _ = image.shape
img_x_min = max(x_min, 0)
img_x_max = min(x_max, w - 1)
img_y_min = max(y_min, 0)
img_y_max = min(y_max, h - 1)
crop_image = np.zeros((new_image_size, new_image_size, 3),
dtype=np.float32)
crop_image[max(0, -y_min):max(0, -y_min) + img_y_max - img_y_min + 1, \
max(0, -x_min):max(0, -x_min) + img_x_max - img_x_min + 1, :] \
= image[img_y_min:img_y_max + 1, img_x_min:img_x_max +1, :]
data_image[frame_id, :, :, :] = scipy.misc.imresize(
crop_image, [image_size, image_size])
#draw_2d_joints_surreal(data_image[frame_id, :, :, :], data_J_2d[frame_id, :, :].astype('int32'), name='/home/local/tmp/new/src'+str(t)+str(frame_id)+'.jpg')
seg_float = output[frame_id]['seg'].astype(np.float32)
crop_seg = np.zeros((new_image_size, new_image_size, 3),
dtype=np.float32)
crop_seg[max(0, -y_min):max(0, -y_min) + img_y_max - img_y_min + 1, \
max(0, -x_min):max(0, -x_min) + img_x_max - img_x_min + 1] \
= np.expand_dims(seg_float[img_y_min:img_y_max + 1, img_x_min:img_x_max +1], 2)
seg = scipy.misc.imresize(crop_seg, [image_size, image_size])
seg[seg < 0.5] = 0
seg[seg >= 0.5] = 1
data_seg[frame_id, :, :] = seg[:, :, 0]
data_chamfer[frame_id, :, :], _, _ = get_chamfer(
seg[:, :, 0], chamfer_scale)
return data_pose, data_T, data_R, data_beta, data_J, data_J_2d, data_J_reconstruct_2d, data_image / 255.0, data_seg, data_f, data_chamfer, data_c, data_gender, data_resize_scale