-
Notifications
You must be signed in to change notification settings - Fork 16
/
layout_refine.py
265 lines (226 loc) · 11.3 KB
/
layout_refine.py
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
130
131
132
133
134
135
136
137
138
139
140
141
142
143
144
145
146
147
148
149
150
151
152
153
154
155
156
157
158
159
160
161
162
163
164
165
166
167
168
169
170
171
172
173
174
175
176
177
178
179
180
181
182
183
184
185
186
187
188
189
190
191
192
193
194
195
196
197
198
199
200
201
202
203
204
205
206
207
208
209
210
211
212
213
214
215
216
217
218
219
220
221
222
223
224
225
226
227
228
229
230
231
232
233
234
235
236
237
238
239
240
241
242
243
244
245
246
247
248
249
250
251
252
253
254
255
256
257
258
259
260
261
262
263
264
# -*- encoding: utf-8 -*-
'''
@File : layout_refine.py
@Time : 2023/01/05 16:18:00
@Author : Chuhao Jin
@Email : [email protected]
'''
# here put the import lib
import torch
import cv2
import os
import numpy as np
from tqdm import tqdm
from utils.anchor_utils import get_text_region
from model.layout_model import BBoxesRegModel
STD_WIDTH, STD_HEIGHT = 300, 400
channel_deep = 64
MIN_VALUE = -999999
MAX_BBOX_NUM = 32
layout_model = BBoxesRegModel(channel_deep = channel_deep)
ckpt_path = "./checkpoint/0.20484_Cascading_128_uniform_big.pth"
layout_model.load_state_dict(torch.load(ckpt_path).module.state_dict())
layout_model.cuda()
layout_model.eval()
# Given the bounding box of a layout, return the layout binary mask.
def get_bbox_mask(bbox, data_len):
mask = np.zeros((1, STD_HEIGHT, STD_WIDTH), dtype = np.uint8)
regions = np.array([[[obj[0], obj[1]],
[obj[2], obj[1]],
[obj[2], obj[3]],
[obj[0], obj[3]]] for obj in bbox[0][:min(MAX_BBOX_NUM, data_len)]], dtype = np.int32)
cv2.fillPoly(mask[0], regions, 1, 1)
return mask
#
def get_batch_text_region(distrib_mask, bbox_size_list, img_size):
temp_mask = distrib_mask.copy().astype(np.float64)
bbox_pos = np.zeros((1, MAX_BBOX_NUM, 4))
pixel_gap = STD_HEIGHT // 10
for i, bbox_size in enumerate(bbox_size_list):
scaled_text_width = int(bbox_size[0] / img_size[0] * STD_WIDTH)
scaled_text_height = int(bbox_size[1] / img_size[1] * STD_HEIGHT)
text_bboxes, text_scores = get_text_region(temp_mask, (scaled_text_width, scaled_text_height), top_n=1)
text_regions = np.array([[[obj[0] - pixel_gap, obj[1] - pixel_gap],
[obj[2]+pixel_gap, obj[1] - pixel_gap],
[obj[2]+pixel_gap, obj[3]+pixel_gap],
[obj[0] - pixel_gap, obj[3]+pixel_gap]]
for obj in text_bboxes])
cv2.fillPoly(temp_mask, text_regions, -10)
bbox_pos[0, i] = text_bboxes[0]
return bbox_pos
# used to deoverlap.
def overlap_detection(bboxes, box_id = -1):
""" We resize both tensors to [A,B,2] without new malloc:
[A,2] -> [A,1,2] -> [A,B,2]
[B,2] -> [1,B,2] -> [A,B,2]
Then we compute the area of intersect between box_a and box_b.
Args:
box_a: (tensor) bounding boxes, Shape: [A,4].
box_b: (tensor) bounding boxes, Shape: [B,4].
Return:
(tensor) intersection area, Shape: [A,B].
"""
A = bboxes.size(0)
B = bboxes.size(0)
max_xy = torch.min(bboxes[:, 2:].unsqueeze(1).expand(A, B, 2),
bboxes[:, 2:].unsqueeze(0).expand(A, B, 2))
min_xy = torch.max(bboxes[:, :2].unsqueeze(1).expand(A, B, 2),
bboxes[:, :2].unsqueeze(0).expand(A, B, 2))
inter = torch.clamp((max_xy - min_xy), min=0)
res = inter[:, :, 0] * inter[:, :, 1]
diag = torch.diag(res)
res_diag = torch.diag_embed(diag)
res = res - res_diag
if box_id == -1:
if res.max() != 0:
return True
else:
return False
if res[box_id].max() != 0:
overlap_id = torch.argmax(res[box_id])
return True, overlap_id
else:
return False, -1
# Given the current layout, predict the refine layout.
def get_next_bbox(cur_bbox, shift, data_len, x_step, y_step, bbox_size, order):
move_pixel = 10
# shift[np.abs(shift)<0.1] = 0
if np.abs(shift).sum() == 0:
return cur_bbox, True
cur_bbox = cur_bbox * np.array((STD_WIDTH, STD_HEIGHT, STD_WIDTH, STD_HEIGHT))
cur_bbox = cur_bbox.astype(np.int32)
shift_scaled = shift * np.array((x_step, y_step))
next_bbox = np.zeros((1, MAX_BBOX_NUM, 4), dtype = np.int32)
next_bbox[0, :data_len] = cur_bbox[0, :data_len]
# print(bbox_size)
for i in range(data_len):
next_bbox[0, i, :2] = cur_bbox[0, i, :2] + shift_scaled[0,i]
next_bbox[0, i, 2:] = next_bbox[0, i, :2] + bbox_size[i]
isoverlap, overlap_id = overlap_detection(torch.tensor(next_bbox[0]), box_id = i)
if isoverlap:
next_bbox[0, i, :2] = cur_bbox[0, i, :2]
next_bbox[0, i, 0] += shift_scaled[0, i, 0]
next_bbox[0, i, 0] = max(min(STD_WIDTH - bbox_size[i, 0], next_bbox[0, i, 0]), 0)
if next_bbox[0, i, 1] < next_bbox[0, overlap_id, 1]:
if next_bbox[0, i, 1] > move_pixel:
next_bbox[0, i, 1] -= move_pixel
else:
next_bbox[0, overlap_id, 1] += move_pixel
next_bbox[0, overlap_id, 3] += move_pixel
elif next_bbox[0, i, 1] >= next_bbox[0, overlap_id, 1]:
if next_bbox[0, i, 1] <=STD_HEIGHT - move_pixel - bbox_size[i, 0]:
next_bbox[0, i, 1] += move_pixel
else:
next_bbox[0, overlap_id, 1] -= move_pixel
next_bbox[0, overlap_id, 3] -= move_pixel
next_bbox[0, i, 2:] = next_bbox[0, i, :2] + bbox_size[i]
next_bbox[0, data_len:] = MIN_VALUE
bbox_argsort = np.argsort(-next_bbox[0, :, 1])
next_bbox[0] = next_bbox[0, bbox_argsort]
bbox_size = bbox_size[bbox_argsort]
order = order[bbox_argsort]
next_bbox[0, data_len:] = 0
return next_bbox, bbox_size, order
# refine the layout iteratively.
def get_refine_bboxes(initial_data, iteration_rounds = 10):
len_info, bbox_mask, shifted_bbox = initial_data["len_info"], initial_data["shifted_mask"], initial_data["shifted_bbox"]
bbox_distrib_map, smooth_region_mask = initial_data["bbox_distrib_map"], initial_data["smooth_region_mask"]
data_len = min(len_info, len_info)
shifted_bbox[0, len_info: ] = MIN_VALUE
sort_order = np.argsort(-shifted_bbox[0, :, 1])
shifted_bbox[0, :, ] = shifted_bbox[0, sort_order]
shifted_bbox[0, len_info: ] = 0
init_bbox_size = shifted_bbox[0, :, 2:] - shifted_bbox[0, :, :2]
len_info = torch.tensor(len_info).unsqueeze(0)
shifted_bbox = torch.tensor(shifted_bbox) / torch.tensor((STD_WIDTH, STD_HEIGHT, STD_WIDTH, STD_HEIGHT))
shifted_bbox = shifted_bbox.float().cuda()
bbox_distrib_map = torch.tensor(bbox_distrib_map).float().unsqueeze(0).unsqueeze(0).cuda()
bbox_mask = torch.tensor(bbox_mask).float().cuda()
smooth_region_mask = torch.tensor(smooth_region_mask).float().cuda()
len_info = len_info.cuda()
order = np.zeros((MAX_BBOX_NUM, ), dtype = np.int32)
order[:len_info] = sort_order[:len_info]
for _ in range(iteration_rounds):
with torch.no_grad():
shifted_pred = layout_model.forward(len_info, bbox_mask, shifted_bbox, None,
bbox_distrib_map, smooth_region_mask, None, inference = True)
next_bbox, init_bbox_size, order = get_next_bbox(shifted_bbox.clone().cpu().numpy(),
shifted_pred.clone().cpu().numpy(),
data_len, x_step = STD_WIDTH // 10, y_step = STD_HEIGHT // 10,
bbox_size = init_bbox_size, order = order)
next_mask = get_bbox_mask(next_bbox, data_len)
shifted_bbox = torch.tensor(next_bbox) / torch.tensor((STD_WIDTH, STD_HEIGHT, STD_WIDTH, STD_HEIGHT))
shifted_bbox = shifted_bbox.cuda()
shifted_mask = torch.tensor(next_mask).cuda()
return next_bbox[0], init_bbox_size[0], order
if __name__ == "__main__":
from utils.font_utils import PutText2Image
from model.distrib_model import LayoutsDistribModel
from layout_distribution_predict import smooth_region_dectection, get_distrib_mask
# A example of layout text.
sentences = [
("Hello, World", 50),
("POSTER", 35),
("LAYOUT-LAYOUT", 35)]
# A example font file.
font_file = "./font_files/test_font.TTF"
text_color = (255, 255 ,255)
# A example background image.
img_path = "./bk_image_folder/3AanCrYzXN0.png"
# The iteration round of layout refine model.
iteration_rounds = 30
ft_center = PutText2Image(font_file)
bbox_size_array = np.zeros((len(sentences), 2))
img = cv2.imread(img_path)
width, height = img.shape[1], img.shape[0]
img_size = (width, height)
data_len = len(sentences)
smooth_region_mask, regions, saliency_map = smooth_region_dectection(img)
bbox_distrib_map = get_distrib_mask(smooth_region_mask)
for i, text_info in enumerate(sentences):
bbox_size_array[i] = ft_center.get_text_bbox_size(text=text_info[0], text_size=text_info[1])
# initial layout, sampled by the maximum probability above the bbox_distrib_map.
initial_bboxes = get_batch_text_region(bbox_distrib_map, bbox_size_array, img_size)
print("initial_bboxes:", initial_bboxes.shape)
initial_bbox_mask = get_bbox_mask(initial_bboxes, data_len)
# The data used to refine the layout.
initial_data = {"len_info": data_len,
"shifted_mask": initial_bbox_mask.copy(),
"shifted_bbox": initial_bboxes.copy(),
"bbox_distrib_map": bbox_distrib_map.copy(),
"smooth_region_mask": smooth_region_mask.copy()}
# The refined layout.
refined_bboxes, refined_bbox_size, order = get_refine_bboxes(initial_data, iteration_rounds)
print("refine_bboxes:", refined_bboxes.shape)
refined_bbox_mask = get_bbox_mask(refined_bboxes[None, :], data_len)
# scale the layout to the image size.
refined_bboxes[:, (0, 2)] = refined_bboxes[:, (0, 2)] / STD_WIDTH * width
refined_bboxes[:, (1, 3)] = refined_bboxes[:, (1, 3)] / STD_HEIGHT * height
# show (salicy_map, predicted_layout_distribution, initial_layout) in a figure.
add_sal_mask = np.zeros((height, width, 3))
initial_bbox_mask = cv2.resize(initial_bbox_mask[0], (width, height))
smooth_region_mask = cv2.resize(smooth_region_mask[0], (width, height))
add_sal_mask[:, :, 2] = initial_bbox_mask / initial_bbox_mask.max()
bbox_distrib_map = cv2.resize(bbox_distrib_map, (width, height))
add_sal_mask[:, :, 1] = bbox_distrib_map / bbox_distrib_map.max()
add_sal_mask[:, :, 0] = saliency_map / saliency_map.max()
add_sal_mask = cv2.resize(add_sal_mask, (width, height))
cv2.imwrite("./display/initial_layout.jpg", add_sal_mask * 255)
# show (salicy_map, predicted_layout_distribution, refined_layout) in a figure.
add_sal_mask = np.zeros((height, width, 3))
refined_bbox_mask = cv2.resize(refined_bbox_mask[0], (width, height))
# smooth_region_mask = cv2.resize(smooth_region_mask[0], (width, height))
add_sal_mask[:, :, 2] = refined_bbox_mask / refined_bbox_mask.max()
bbox_distrib_map = cv2.resize(bbox_distrib_map, (width, height))
add_sal_mask[:, :, 1] = bbox_distrib_map / bbox_distrib_map.max()
add_sal_mask[:, :, 0] = saliency_map / saliency_map.max()
add_sal_mask = cv2.resize(add_sal_mask, (width, height))
cv2.imwrite("./display/refined_layout.jpg", add_sal_mask * 255)
# save the final poster picture.
bk_img = img.copy()
bk_img = bk_img.astype(np.uint8)
for j in range(len(sentences)):
text_position = (refined_bboxes[j][0], refined_bboxes[j][1])
text, text_size = sentences[order[j]][0], sentences[order[j]][1]
bk_img = ft_center.draw_text(bk_img, text_position, text, text_size, text_color)
cv2.imwrite("./display/poster.jpg", bk_img)