Spaces:
Running
on
Zero
Running
on
Zero
Upload run_facer_segmentation.py
Browse files
src/pixel3dmm/run_facer_segmentation.py
ADDED
@@ -0,0 +1,208 @@
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
1 |
+
import os
|
2 |
+
import sys
|
3 |
+
import traceback
|
4 |
+
|
5 |
+
from math import ceil
|
6 |
+
|
7 |
+
import PIL.Image
|
8 |
+
import torch
|
9 |
+
import distinctipy
|
10 |
+
import matplotlib.pyplot as plt
|
11 |
+
from PIL import Image
|
12 |
+
import numpy as np
|
13 |
+
import tyro
|
14 |
+
import facer
|
15 |
+
|
16 |
+
from pixel3dmm import env_paths
|
17 |
+
|
18 |
+
colors = distinctipy.get_colors(22, rng=0)
|
19 |
+
|
20 |
+
|
21 |
+
def viz_results(img, seq_classes, n_classes, suppress_plot = False):
|
22 |
+
|
23 |
+
seg_img = np.zeros([img.shape[-2], img.shape[-1], 3])
|
24 |
+
#distinctipy.color_swatch(colors)
|
25 |
+
bad_indices = [
|
26 |
+
0, # background,
|
27 |
+
1, # neck
|
28 |
+
# 2, skin
|
29 |
+
3, # cloth
|
30 |
+
4, # ear_r (images-space r)
|
31 |
+
5, # ear_l
|
32 |
+
# 6 brow_r
|
33 |
+
# 7 brow_l
|
34 |
+
# 8, # eye_r
|
35 |
+
# 9, # eye_l
|
36 |
+
# 10 noise
|
37 |
+
# 11 mouth
|
38 |
+
# 12 lower_lip
|
39 |
+
# 13 upper_lip
|
40 |
+
14, # hair,
|
41 |
+
# 15, glasses
|
42 |
+
16, # ??
|
43 |
+
17, # earring_r
|
44 |
+
18, # ?
|
45 |
+
]
|
46 |
+
bad_indices = []
|
47 |
+
|
48 |
+
for i in range(n_classes):
|
49 |
+
if i not in bad_indices:
|
50 |
+
seg_img[seq_classes[0, :, :] == i] = np.array(colors[i])*255
|
51 |
+
|
52 |
+
if not suppress_plot:
|
53 |
+
plt.imshow(seg_img.astype(np.uint(8)))
|
54 |
+
plt.show()
|
55 |
+
return Image.fromarray(seg_img.astype(np.uint8))
|
56 |
+
|
57 |
+
def get_color_seg(img, seq_classes, n_classes):
|
58 |
+
|
59 |
+
seg_img = np.zeros([img.shape[-2], img.shape[-1], 3])
|
60 |
+
colors = distinctipy.get_colors(n_classes+1, rng=0)
|
61 |
+
#distinctipy.color_swatch(colors)
|
62 |
+
bad_indices = [
|
63 |
+
0, # background,
|
64 |
+
1, # neck
|
65 |
+
# 2, skin
|
66 |
+
3, # cloth
|
67 |
+
4, # ear_r (images-space r)
|
68 |
+
5, # ear_l
|
69 |
+
# 6 brow_r
|
70 |
+
# 7 brow_l
|
71 |
+
# 8, # eye_r
|
72 |
+
# 9, # eye_l
|
73 |
+
# 10 noise
|
74 |
+
# 11 mouth
|
75 |
+
# 12 lower_lip
|
76 |
+
# 13 upper_lip
|
77 |
+
14, # hair,
|
78 |
+
# 15, glasses
|
79 |
+
16, # ??
|
80 |
+
17, # earring_r
|
81 |
+
18, # ?
|
82 |
+
]
|
83 |
+
|
84 |
+
for i in range(n_classes):
|
85 |
+
if i not in bad_indices:
|
86 |
+
seg_img[seq_classes[0, :, :] == i] = np.array(colors[i])*255
|
87 |
+
|
88 |
+
|
89 |
+
return Image.fromarray(seg_img.astype(np.uint8))
|
90 |
+
|
91 |
+
|
92 |
+
def crop_gt_img(img, seq_classes, n_classes):
|
93 |
+
|
94 |
+
seg_img = np.zeros([img.shape[-2], img.shape[-1], 3])
|
95 |
+
colors = distinctipy.get_colors(n_classes+1, rng=0)
|
96 |
+
#distinctipy.color_swatch(colors)
|
97 |
+
bad_indices = [
|
98 |
+
0, # background,
|
99 |
+
1, # neck
|
100 |
+
# 2, skin
|
101 |
+
3, # cloth
|
102 |
+
4, #ear_r (images-space r)
|
103 |
+
5, #ear_l
|
104 |
+
# 6 brow_r
|
105 |
+
# 7 brow_l
|
106 |
+
#8, # eye_r
|
107 |
+
#9, # eye_l
|
108 |
+
# 10 noise
|
109 |
+
# 11 mouth
|
110 |
+
# 12 lower_lip
|
111 |
+
# 13 upper_lip
|
112 |
+
14, # hair,
|
113 |
+
# 15, glasses
|
114 |
+
16, # ??
|
115 |
+
17, # earring_r
|
116 |
+
18, # ?
|
117 |
+
]
|
118 |
+
|
119 |
+
for i in range(n_classes):
|
120 |
+
if i in bad_indices:
|
121 |
+
img[seq_classes[0, :, :] == i] = 0
|
122 |
+
|
123 |
+
|
124 |
+
#plt.imshow(img.astype(np.uint(8)))
|
125 |
+
#plt.show()
|
126 |
+
return img.astype(np.uint8)
|
127 |
+
|
128 |
+
|
129 |
+
def segment(video_name : str, face_detector, face_parser):
|
130 |
+
|
131 |
+
|
132 |
+
out = f'{env_paths.PREPROCESSED_DATA}/{video_name}'
|
133 |
+
out_seg = f'{out}/seg_og/'
|
134 |
+
out_seg_annot = f'{out}/seg_non_crop_annotations/'
|
135 |
+
os.makedirs(out_seg, exist_ok=True)
|
136 |
+
os.makedirs(out_seg_annot, exist_ok=True)
|
137 |
+
folder = f'{out}/cropped/' # '/home/giebenhain/GTA/data_kinect/color/'
|
138 |
+
|
139 |
+
|
140 |
+
|
141 |
+
|
142 |
+
|
143 |
+
frames = [f for f in os.listdir(folder) if f.endswith('.png') or f.endswith('.jpg')]
|
144 |
+
|
145 |
+
frames.sort()
|
146 |
+
|
147 |
+
if len(os.listdir(out_seg)) == len(frames):
|
148 |
+
print(f'''
|
149 |
+
<<<<<<<< ALREADY COMPLETED SEGMENTATION FOR {video_name}, SKIPPING >>>>>>>>
|
150 |
+
''')
|
151 |
+
return
|
152 |
+
|
153 |
+
#for file in frames:
|
154 |
+
batch_size = 1
|
155 |
+
|
156 |
+
for i in range(len(frames)//batch_size):
|
157 |
+
image_stack = []
|
158 |
+
frame_stack = []
|
159 |
+
original_shapes = []
|
160 |
+
for j in range(batch_size):
|
161 |
+
file = frames[i * batch_size + j]
|
162 |
+
|
163 |
+
if os.path.exists(f'{out_seg_annot}/color_{file}.png'):
|
164 |
+
print('DONE')
|
165 |
+
continue
|
166 |
+
img = Image.open(f'{folder}/{file}')#.resize((512, 512))
|
167 |
+
|
168 |
+
og_size = img.size
|
169 |
+
|
170 |
+
image = facer.hwc2bchw(torch.from_numpy(np.array(img)[..., :3])).to(device="cuda") # image: 1 x 3 x h x w
|
171 |
+
image_stack.append(image)
|
172 |
+
frame_stack.append(file[:-4])
|
173 |
+
|
174 |
+
for batch_idx in range(ceil(len(image_stack)/batch_size)):
|
175 |
+
image_batch = torch.cat(image_stack[batch_idx*batch_size:(batch_idx+1)*batch_size], dim=0)
|
176 |
+
frame_idx_batch = frame_stack[batch_idx*batch_size:(batch_idx+1)*batch_size]
|
177 |
+
og_shape_batch = original_shapes[batch_idx*batch_size:(batch_idx+1)*batch_size]
|
178 |
+
|
179 |
+
#if True:
|
180 |
+
try:
|
181 |
+
with torch.inference_mode():
|
182 |
+
faces = face_detector(image_batch)
|
183 |
+
torch.cuda.empty_cache()
|
184 |
+
faces = face_parser(image_batch, faces, bbox_scale_factor=1.25)
|
185 |
+
torch.cuda.empty_cache()
|
186 |
+
|
187 |
+
seg_logits = faces['seg']['logits']
|
188 |
+
back_ground = torch.all(seg_logits == 0, dim=1, keepdim=True).detach().squeeze(1).cpu().numpy()
|
189 |
+
seg_probs = seg_logits.softmax(dim=1) # nfaces x nclasses x h x w
|
190 |
+
seg_classes = seg_probs.argmax(dim=1).detach().cpu().numpy().astype(np.uint8)
|
191 |
+
seg_classes[back_ground] = seg_probs.shape[1] + 1
|
192 |
+
|
193 |
+
|
194 |
+
for _iidx in range(seg_probs.shape[0]):
|
195 |
+
frame = frame_idx_batch[_iidx]
|
196 |
+
iidx = faces['image_ids'][_iidx].item()
|
197 |
+
try:
|
198 |
+
I_color = viz_results(image_batch[iidx:iidx+1], seq_classes=seg_classes[_iidx:_iidx+1], n_classes=seg_probs.shape[1] + 1, suppress_plot=True)
|
199 |
+
I_color.save(f'{out_seg_annot}/color_{frame}.png')
|
200 |
+
except Exception as ex:
|
201 |
+
pass
|
202 |
+
I = Image.fromarray(seg_classes[_iidx])
|
203 |
+
I.save(f'{out_seg}/{frame}.png')
|
204 |
+
torch.cuda.empty_cache()
|
205 |
+
except Exception as exx:
|
206 |
+
traceback.print_exc()
|
207 |
+
continue
|
208 |
+
|