forked from yijingru/Vertebra-Landmark-Detection
-
Notifications
You must be signed in to change notification settings - Fork 0
/
pre_proc.py
146 lines (127 loc) · 5.23 KB
/
pre_proc.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
import cv2
import torch
from draw_gaussian import *
import transform
import math
def processing_test(image, input_h, input_w):
image = cv2.resize(image, (input_w, input_h))
out_image = image.astype(np.float32) / 255.
out_image = out_image - 0.5
out_image = out_image.transpose(2, 0, 1).reshape(1, 3, input_h, input_w)
out_image = torch.from_numpy(out_image)
return out_image
def draw_spinal(pts, out_image):
colors = [(0, 0, 255), (0, 255, 255), (255, 0, 255), (0, 255, 0)]
for i in range(4):
cv2.circle(out_image, (int(pts[i, 0]), int(pts[i, 1])), 3, colors[i], 1, 1)
cv2.putText(out_image, '{}'.format(i+1), (int(pts[i, 0]), int(pts[i, 1])),
cv2.FONT_HERSHEY_SIMPLEX, 0.3, (0,0,0),1,1)
for i,j in zip([0,1,2,3], [1,2,3,0]):
cv2.line(out_image,
(int(pts[i, 0]), int(pts[i, 1])),
(int(pts[j, 0]), int(pts[j, 1])),
color=colors[i], thickness=1, lineType=1)
return out_image
def rearrange_pts(pts):
# rearrange left right sequence
boxes = []
centers = []
for k in range(0, len(pts), 4):
pts_4 = pts[k:k+4,:]
x_inds = np.argsort(pts_4[:, 0])
pt_l = np.asarray(pts_4[x_inds[:2], :])
pt_r = np.asarray(pts_4[x_inds[2:], :])
y_inds_l = np.argsort(pt_l[:,1])
y_inds_r = np.argsort(pt_r[:,1])
tl = pt_l[y_inds_l[0], :]
bl = pt_l[y_inds_l[1], :]
tr = pt_r[y_inds_r[0], :]
br = pt_r[y_inds_r[1], :]
# boxes.append([tl, tr, bl, br])
boxes.append(tl)
boxes.append(tr)
boxes.append(bl)
boxes.append(br)
centers.append(np.mean(pts_4, axis=0))
bboxes = np.asarray(boxes, np.float32)
# rearrange top to bottom sequence
centers = np.asarray(centers, np.float32)
sort_tb = np.argsort(centers[:,1])
new_bboxes = []
for sort_i in sort_tb:
new_bboxes.append(bboxes[4*sort_i, :])
new_bboxes.append(bboxes[4*sort_i+1, :])
new_bboxes.append(bboxes[4*sort_i+2, :])
new_bboxes.append(bboxes[4*sort_i+3, :])
new_bboxes = np.asarray(new_bboxes, np.float32)
return new_bboxes
def generate_ground_truth(image,
pts_2,
image_h,
image_w,
img_id):
hm = np.zeros((1, image_h, image_w), dtype=np.float32)
wh = np.zeros((17, 2*4), dtype=np.float32)
reg = np.zeros((17, 2), dtype=np.float32)
ind = np.zeros((17), dtype=np.int64)
reg_mask = np.zeros((17), dtype=np.uint8)
if pts_2[:,0].max()>image_w:
print('w is big', pts_2[:,0].max())
if pts_2[:,1].max()>image_h:
print('h is big', pts_2[:,1].max())
if pts_2.shape[0]!=68:
print('ATTENTION!! image {} pts does not equal to 68!!! '.format(img_id))
for k in range(17):
pts = pts_2[4*k:4*k+4,:]
bbox_h = np.mean([np.sqrt(np.sum((pts[0,:]-pts[2,:])**2)),
np.sqrt(np.sum((pts[1,:]-pts[3,:])**2))])
bbox_w = np.mean([np.sqrt(np.sum((pts[0,:]-pts[1,:])**2)),
np.sqrt(np.sum((pts[2,:]-pts[3,:])**2))])
cen_x, cen_y = np.mean(pts, axis=0)
ct = np.asarray([cen_x, cen_y], dtype=np.float32)
ct_int = ct.astype(np.int32)
radius = gaussian_radius((math.ceil(bbox_h), math.ceil(bbox_w)))
radius = max(0, int(radius))
draw_umich_gaussian(hm[0,:,:], ct_int, radius=radius)
ind[k] = ct_int[1] * image_w + ct_int[0]
reg[k] = ct - ct_int
reg_mask[k] = 1
for i in range(4):
wh[k,2*i:2*i+2] = ct-pts[i,:]
ret = {'input': image,
'hm': hm,
'ind': ind,
'reg': reg,
'wh': wh,
'reg_mask': reg_mask,
}
return ret
# def filter_pts(pts, w, h):
# pts_new = []
# for pt in pts:
# if any(pt) < 0 or pt[0] > w - 1 or pt[1] > h - 1:
# continue
# else:
# pts_new.append(pt)
# return np.asarray(pts_new, np.float32)
def processing_train(image, pts, image_h, image_w, down_ratio, aug_label, img_id):
# filter pts ----------------------------------------------------
h,w,c = image.shape
# pts = filter_pts(pts, w, h)
# ---------------------------------------------------------------
data_aug = {'train': transform.Compose([transform.ConvertImgFloat(),
transform.PhotometricDistort(),
transform.Expand(max_scale=1.5, mean=(0, 0, 0)),
transform.RandomMirror_w(),
transform.Resize(h=image_h, w=image_w)]),
'val': transform.Compose([transform.ConvertImgFloat(),
transform.Resize(h=image_h, w=image_w)])}
if aug_label:
out_image, pts = data_aug['train'](image.copy(), pts)
else:
out_image, pts = data_aug['val'](image.copy(), pts)
out_image = np.clip(out_image, a_min=0., a_max=255.)
out_image = np.transpose(out_image / 255. - 0.5, (2,0,1))
pts = rearrange_pts(pts)
pts2 = transform.rescale_pts(pts, down_ratio=down_ratio)
return np.asarray(out_image, np.float32), pts2