-
Notifications
You must be signed in to change notification settings - Fork 1
/
inference.py
325 lines (244 loc) · 12.5 KB
/
inference.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
265
266
267
268
269
270
271
272
273
274
275
276
277
278
279
280
281
282
283
284
285
286
287
288
289
290
291
292
293
294
295
296
297
298
299
300
301
302
303
304
305
306
307
308
309
310
311
312
313
314
315
316
317
318
319
320
321
322
323
324
325
import argparse
import os
import json
from collections import OrderedDict
from glob import glob
from tqdm import tqdm
import cv2
import numpy as np
import torch
from torch.utils.data import DataLoader
from torchvision import transforms as T
import torchvision.ops as ops
from models import U2NETP
from data_loader import InferenceDataset
# from models.yolov4 import Darknet
# from models.yolov4.torch_utils import do_detect
from PyTorch_YOLOv4.utils.general import non_max_suppression,box_iou
from PyTorch_YOLOv4.models.models import *
# from non_max import filter_predictions, non_max_suppression
def findBboxes(label, original_shape, current_shape, det_size=960):
H,W = original_shape
_H,_W = current_shape
contours = cv2.findContours(label, cv2.RETR_EXTERNAL, cv2.CHAIN_APPROX_SIMPLE)
contours = contours[0] if len(contours) == 2 else contours[1]
bboxes = []
for cntr in contours:
x,y,w,h = cv2.boundingRect(cntr)
xmin, ymin = (x/_W)*W, (y/_H)*H
xmax, ymax = ((x+w)/_W)*W, ((y+h)/_H)*H
bbox = [int(x) for x in [xmin,ymin,xmax,ymax]]
bboxes.append(bbox)
return bboxes
def isInsidePoint(bbox, point):
xmin, ymin, xmax, ymax = bbox
x, y = point
if xmin<=x<=xmax and ymin<=y<=ymax:
return True
def isInsideBbox(inner_bbox, outer_bbox):
xmin,ymin,xmax,ymax = inner_bbox
p1, p2, p3, p4 = (xmin,ymin), (xmax,ymin), (xmax, ymax), (xmin, ymax)
return all([isInsidePoint(outer_bbox, point) for point in [p1,p2,p3,p4]])
def getDetectionBbox(bbox, max_H, max_W, det_size=960):
xmin, ymin, xmax, ymax = bbox
xc = (xmax+xmin)//2
yc = (ymin+ymax)//2
xmin = max(xc - det_size//2, 0)
ymin = max(yc - det_size//2, 0)
xmax = min(xmin+det_size, max_W)
ymax = min(ymin+det_size, max_H)
return [xmin,ymin,xmax,ymax]
def getDetectionBboxesAll(bboxes, max_H, max_W, det_size=960):
det_bboxes = []
for bbox in bboxes:
det_bboxes.append(getDetectionBbox(bbox, max_H, max_W, det_size=det_size))
return det_bboxes
def getDetectionBboxesNaive(bboxes, max_H, max_W, det_size=960):
det_bboxes = []
for bbox in bboxes:
if any([isInsideBbox(bbox, det_bbox) for det_bbox in det_bboxes]):
continue
det_bboxes.append(getDetectionBbox(bbox, max_H, max_W, det_size=det_size))
return det_bboxes
def getDetectionBboxesSorted(bboxes, max_H, max_W, det_size=960):
_det_bboxes = [getDetectionBbox(bbox, max_H, max_W, det_size=det_size) for bbox in bboxes]
hits = {i: 0 for i in range(len(_det_bboxes))}
for i, det_bbox in enumerate(_det_bboxes):
hits[i]+=sum([isInsideBbox(bbox, det_bbox) for bbox in bboxes])
# print(hits)
if all([x==1 for x in hits.values()]):
return _det_bboxes
elif any([x==len(bboxes) for x in hits.values()]):
fnd = list(hits.keys())[list(hits.values()).index(len(bboxes))]
# print(fnd)
return [_det_bboxes[fnd]]
else:
hits = dict(sorted(hits.items(), key=lambda item: item[1], reverse=True))
bboxes = [bboxes[i] for i in hits.keys()]
return getDetectionBboxesNaive(bboxes, max_H, max_W, det_size=det_size)
def getDetectionBboxes(bboxes, max_H, max_W, det_size=960, bbox_type='naive'):
if bbox_type == 'naive':
return getDetectionBboxesNaive(bboxes, max_H, max_W, det_size=det_size)
elif bbox_type == 'all':
return getDetectionBboxesAll(bboxes, max_H, max_W, det_size=det_size)
elif bbox_type == 'sorted':
return getDetectionBboxesSorted(bboxes, max_H, max_W, det_size=det_size)
else:
raise NotImplementedError
def getSlidingWindowBBoxes(max_H, max_W, overlap_frac, det_size=960):
n_w = math.ceil(W_orig / (det_size*(1-overlap_frac)))
n_h = math.ceil(H_orig / (det_size*(1-overlap_frac)))
XS = [det_size*(1-overlap_frac)*n for n in range(n_w)]
YS = [det_size*(1-overlap_frac)*n for n in range(n_h)]
bboxes = []
for xmin in XS:
for ymin in YS:
xmin,ymin = [int(_) for _ in [xmin,ymin]]
xmax,ymax = min(xmin+det_size, max_W), min(ymin+det_size, max_H)
bboxes.append([xmin, ymin, xmax, ymax])
return bboxes
def NMS(prediction, iou_thres, redundant=True, merge=False, max_det=300, agnostic=False):
# https://github.com/WongKinYiu/PyTorch_YOLOv4/blob/master/utils/general.py
# output = [torch.zeros(0, 6)] * len(prediction)
x = prediction.clone()
# Batched NMS
boxes, scores, c = x[:, :4], x[:, 4], x[:, 5] * (0 if agnostic else 1)
i = ops.batched_nms(boxes, scores, c, iou_thres)
# i = torch.ops.torchvision.nms(boxes, scores, iou_thres)
if i.shape[0] > max_det: # limit detections
i = i[:max_det]
if merge and (1 < x.shape[0] < 3E3): # Merge NMS (boxes merged using weighted mean)
# update boxes as boxes(i,4) = weights(i,n) * boxes(n,4)
# iou = bbox_iou(torch.unsqueeze(boxes[i], 0), boxes,DIoU=True) > iou_thres
iou = box_iou(boxes[i], boxes) > iou_thres # iou matrix
weights = iou * scores[None] # box weights
x[i, :4] = torch.mm(weights, x[:, :4]).float() / weights.sum(1, keepdim=True) # merged boxes
if redundant:
i = i[iou.sum(1) > 1] # require redundancy
# output[xi] = x[i]
return x[i] #output
if __name__ == '__main__':
parser = argparse.ArgumentParser()
# ROI
parser.add_argument('--roi_ckpt', type=str, default='trained_models/u2netp.pth')
parser.add_argument('--roi_inf_size', type=int, default=576)
parser.add_argument('--roi_th', type=float, default=0.5)
parser.add_argument('--dilate', default=False, action='store_true')
parser.add_argument('--k_size', type=int, default=7)
parser.add_argument('--iter', type=int, default=2)
parser.add_argument('--bbox_type', type=str, default='naive', choices=['all', 'naive', 'sorted'])
# net_det
parser.add_argument('--det_ckpt', type=str, default='trained_models/mtsd001_best.weights')
parser.add_argument('--det_cfg', type=str, default='trained_models/mtsd001-960.cfg')
parser.add_argument('--det_inf_size', type=int, default=960)
# NMS
parser.add_argument('--conf_thresh', type=float, default=0.005)
parser.add_argument('--iou_thresh', type=float, default=0.45)
parser.add_argument('--second_nms', default=False, action='store_true')
# sliding window
parser.add_argument('--overlap_frac', type=float, default=0.05)
# general
parser.add_argument('--input_files', type=str, default='input_list.txt')
parser.add_argument('--mode', type=str, default='roi', choices=['det', 'roi', 'sw'])
parser.add_argument('--cpu', default=False, action='store_true')
parser.add_argument('--save_results', default=False, action='store_true')
parser.add_argument('--out_dir', type=str, default='detections')
parser.add_argument('--out_json', type=str, default='labels.json')
args = parser.parse_args()
device = torch.device('cuda:0' if torch.cuda.device_count() > 0 and not args.cpu else 'cpu')
if args.save_results:
os.makedirs(args.out_dir, exist_ok=True)
flist = sorted([x.rstrip() for x in open(args.input_files) if os.path.isfile(x.rstrip())])
dataset = InferenceDataset(
img_name_list=flist,
roi_inf_size=args.roi_inf_size)
dataloader = DataLoader(dataset, batch_size=1, shuffle=False, num_workers=4)
if args.mode == 'roi':
net_roi = U2NETP(3,1)
net_roi.load_state_dict(torch.load(args.roi_ckpt, map_location='cpu'))
net_roi.to(device)
net_roi.eval()
net_det = Darknet(args.det_cfg, img_size=(args.det_inf_size, args.det_inf_size)).to(device)
load_darknet_weights(net_det, args.det_ckpt)
net_det.eval()
num_dets = 0
object_id = 0
annotations = []
with torch.no_grad():
for i, data in tqdm(enumerate(dataloader)):
intensor_roi = data['image_roi'].float().to(device)
intensor_det = data['image_det'].float().to(device)
H_orig, W_orig = data['height'].item(), data['width'].item()
original_shape = (H_orig, W_orig)
if args.save_results:
image = cv2.imread(flist[i])
if args.mode == 'det':
tr = T.Resize((args.det_inf_size, args.det_inf_size))
intensor_detection = tr(intensor_det)
inf_out, train_out = net_det(intensor_detection, augment=False)
num_dets+=1
img_output = non_max_suppression(
inf_out,
conf_thres = args.conf_thresh,
iou_thres = args.iou_thresh,
merge = True,
agnostic = False)[0] # batch==1
img_output = img_output.to(device)
img_output[:,:4] *= torch.tensor([W_orig/args.det_inf_size,H_orig/args.det_inf_size,W_orig/args.det_inf_size,H_orig/args.det_inf_size]).to(device)
elif args.mode == 'roi':
d0 = net_roi(intensor_roi)[0]
d0 = d0[0,0,...]
d0 = torch.where(d0>args.roi_th, 1.0, 0.0)
d0 = 255*d0.detach().cpu().numpy().astype(np.uint8)
if args.dilate:
kernel = np.ones((args.k_size,args.k_size), np.uint8)
d0 = cv2.dilate(d0,kernel,iterations = args.iter)
bboxes = findBboxes(d0, original_shape, (args.roi_inf_size, args.roi_inf_size))
bboxes_det = getDetectionBboxes(bboxes, H_orig, W_orig, det_size=args.det_inf_size, bbox_type=args.bbox_type)
elif args.mode == 'sw':
bboxes_det = getSlidingWindowBBoxes(H_orig, W_orig, args.overlap_frac, det_size=args.det_inf_size)
if args.mode == 'sw' or args.mode =='roi':
img_output = None
for bbox_det in bboxes_det:
xmin, ymin, xmax, ymax = bbox_det
intensor_detection = torch.zeros((1, 3, args.det_inf_size, args.det_inf_size)).to(device)
intensor_detection[...,:ymax-ymin,:xmax-xmin] = intensor_det[...,ymin:ymax,xmin:xmax]
H,W = intensor_detection.shape[2:]
inf_out, train_out = net_det(intensor_detection, augment=False)
num_dets+=1
output = non_max_suppression(
inf_out,
conf_thres = args.conf_thresh,
iou_thres = args.iou_thresh,
merge = True,
agnostic = False)[0] # batch==1
output[:,:4] = output[:,:4].to(device) + torch.tensor([xmin,ymin,xmin,ymin]).to(device)
if img_output is None:
img_output = output.clone()
else:
img_output = torch.cat((img_output.to(device), output.to(device)), 0)
if args.second_nms and img_output is not None:
img_output = NMS(img_output, iou_thres=args.iou_thresh, redundant=args.redundant, merge=args.merge, max_det=args.max_det, agnostic=args.agnostic)
if img_output is not None:
for det in img_output:
det = det.cpu().numpy()
_xmin,_ymin,_xmax,_ymax = det[:4]
w,h = _xmax-_xmin, _ymax-_ymin
area = w*h
dct = {
"id": int(object_id),
"image_id": i+1,
"category_id": int(det[-1]),
"bbox" : [float(_) for _ in [_xmin, _ymin, w, h]],
"score": float(det[-2]),
"area": float(area)
}
annotations.append(dct)
object_id+=1
if args.save_results:
image = cv2.rectangle(image, (int(_xmin),int(_ymin)), (int(_xmax),int(_ymax)), (0,0,255), 4)
if args.save_results:
cv2.imwrite('%s/%03d.png' % (args.out_dir, i), image)
with open(args.out_json, 'w', encoding='utf-8') as f:
json.dump(annotations, f, ensure_ascii=False, indent=4)
print('Number of yolov4 detections (%s): %d' % (os.path.basename(args.out_json), num_dets))