Spaces:
Paused
Paused
File size: 5,210 Bytes
a22eb82 ff4c585 a22eb82 |
1 2 3 4 5 6 7 8 9 10 11 12 13 14 15 16 17 18 19 20 21 22 23 24 25 26 27 28 29 30 31 32 33 34 35 36 37 38 39 40 41 42 43 44 45 46 47 48 49 50 51 52 53 54 55 56 57 58 59 60 61 62 63 64 65 66 67 68 69 70 71 72 73 74 75 76 77 78 79 80 81 82 83 84 85 86 87 88 89 90 91 92 93 94 95 96 97 98 99 100 101 102 103 104 105 106 107 108 109 110 111 112 113 114 115 116 117 118 119 120 121 122 123 124 125 126 127 128 129 |
import os
import numpy as np
from PIL import Image
from skimage import io, img_as_float32, transform
import torch
import scipy.io as scio
def get_facerender_data(coeff_path, pic_path, first_coeff_path, audio_path,
batch_size, camera_yaw_list=[0], camera_pitch_list=[0], camera_roll_list=[0],
expression_scale=1.0, still_mode = False):
semantic_radius = 13
video_name = os.path.splitext(os.path.split(coeff_path)[-1])[0]
txt_path = os.path.splitext(coeff_path)[0]
data={}
img1 = Image.open(pic_path)
source_image = np.array(img1)
source_image = img_as_float32(source_image)
source_image = transform.resize(source_image, (256, 256, 3))
source_image = source_image.transpose((2, 0, 1))
source_image_ts = torch.FloatTensor(source_image).unsqueeze(0)
source_image_ts = source_image_ts.repeat(batch_size, 1, 1, 1)
data['source_image'] = source_image_ts
source_semantics_dict = scio.loadmat(first_coeff_path)
source_semantics = source_semantics_dict['coeff_3dmm'][:1,:70] #1 70
source_semantics_new = transform_semantic_1(source_semantics, semantic_radius)
source_semantics_ts = torch.FloatTensor(source_semantics_new).unsqueeze(0)
source_semantics_ts = source_semantics_ts.repeat(batch_size, 1, 1)
data['source_semantics'] = source_semantics_ts
# target
generated_dict = scio.loadmat(coeff_path)
generated_3dmm = generated_dict['coeff_3dmm']
generated_3dmm[:, :64] = generated_3dmm[:, :64] * expression_scale
if still_mode:
generated_3dmm[:, 64:] = np.repeat(source_semantics[:, 64:], generated_3dmm.shape[0], axis=0)
with open(txt_path+'.txt', 'w') as f:
for coeff in generated_3dmm:
for i in coeff:
f.write(str(i)[:7] + ' '+'\t')
f.write('\n')
target_semantics_list = []
frame_num = generated_3dmm.shape[0]
data['frame_num'] = frame_num
for frame_idx in range(frame_num):
target_semantics = transform_semantic_target(generated_3dmm, frame_idx, semantic_radius)
target_semantics_list.append(target_semantics)
remainder = frame_num%batch_size
if remainder!=0:
for _ in range(batch_size-remainder):
target_semantics_list.append(target_semantics)
target_semantics_np = np.array(target_semantics_list) #frame_num 70 semantic_radius*2+1
target_semantics_np = target_semantics_np.reshape(batch_size, -1, target_semantics_np.shape[-2], target_semantics_np.shape[-1])
data['target_semantics_list'] = torch.FloatTensor(target_semantics_np)
data['video_name'] = video_name
data['audio_path'] = audio_path
yaw_c_seq = gen_camera_pose(camera_yaw_list, frame_num, batch_size)
pitch_c_seq = gen_camera_pose(camera_pitch_list, frame_num, batch_size)
roll_c_seq = gen_camera_pose(camera_roll_list, frame_num, batch_size)
data['yaw_c_seq'] = torch.FloatTensor(yaw_c_seq)
data['pitch_c_seq'] = torch.FloatTensor(pitch_c_seq)
data['roll_c_seq'] = torch.FloatTensor(roll_c_seq)
return data
def transform_semantic_1(semantic, semantic_radius):
semantic_list = [semantic for i in range(0, semantic_radius*2+1)]
coeff_3dmm = np.concatenate(semantic_list, 0)
return coeff_3dmm.transpose(1,0)
def transform_semantic_target(coeff_3dmm, frame_index, semantic_radius):
num_frames = coeff_3dmm.shape[0]
seq = list(range(frame_index- semantic_radius, frame_index+ semantic_radius+1))
index = [ min(max(item, 0), num_frames-1) for item in seq ]
coeff_3dmm_g = coeff_3dmm[index, :]
return coeff_3dmm_g.transpose(1,0)
def gen_camera_pose(camera_degree_list, frame_num, batch_size):
new_degree_list = []
if len(camera_degree_list) == 1:
for _ in range(frame_num):
new_degree_list.append(camera_degree_list[0])
remainder = frame_num%batch_size
if remainder!=0:
for _ in range(batch_size-remainder):
new_degree_list.append(new_degree_list[-1])
new_degree_np = np.array(new_degree_list).reshape(batch_size, -1)
return new_degree_np
degree_sum = 0.
for i, degree in enumerate(camera_degree_list[1:]):
degree_sum += abs(degree-camera_degree_list[i])
degree_per_frame = degree_sum/(frame_num-1)
for i, degree in enumerate(camera_degree_list[1:]):
degree_last = camera_degree_list[i]
degree_step = degree_per_frame * abs(degree-degree_last)/(degree-degree_last)
new_degree_list = new_degree_list + list(np.arange(degree_last, degree, degree_step))
if len(new_degree_list) > frame_num:
new_degree_list = new_degree_list[:frame_num]
elif len(new_degree_list) < frame_num:
for _ in range(frame_num-len(new_degree_list)):
new_degree_list.append(new_degree_list[-1])
print(len(new_degree_list))
print(frame_num)
remainder = frame_num%batch_size
if remainder!=0:
for _ in range(batch_size-remainder):
new_degree_list.append(new_degree_list[-1])
new_degree_np = np.array(new_degree_list).reshape(batch_size, -1)
return new_degree_np
|