-
Notifications
You must be signed in to change notification settings - Fork 32
/
demo.py
271 lines (226 loc) · 8.78 KB
/
demo.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
import argparse
import cv2
import os
import time
import numpy as np
import torch
import imageio
from PIL import Image
from dataset.transforms import BaseTransform
from utils.misc import load_weight
from utils.box_ops import rescale_bboxes
from utils.vis_tools import vis_detection
from config import build_dataset_config, build_model_config
from models import build_model
def parse_args():
parser = argparse.ArgumentParser(description='YOWOv2 Demo')
# basic
parser.add_argument('-size', '--img_size', default=224, type=int,
help='the size of input frame')
parser.add_argument('--show', action='store_true', default=False,
help='show the visulization results.')
parser.add_argument('--cuda', action='store_true', default=False,
help='use cuda.')
parser.add_argument('--save_folder', default='det_results/', type=str,
help='Dir to save results')
parser.add_argument('-vs', '--vis_thresh', default=0.3, type=float,
help='threshold for visualization')
parser.add_argument('--video', default='9Y_l9NsnYE0.mp4', type=str,
help='AVA video name.')
parser.add_argument('--gif', action='store_true', default=False,
help='generate gif.')
# class label config
parser.add_argument('-d', '--dataset', default='ava_v2.2',
help='ava_v2.2')
parser.add_argument('--pose', action='store_true', default=False,
help='show 14 action pose of AVA.')
# model
parser.add_argument('-v', '--version', default='yowo_v2_large', type=str,
help='build YOWOv2')
parser.add_argument('--weight', default=None,
type=str, help='Trained state_dict file path to open')
parser.add_argument('-ct', '--conf_thresh', default=0.1, type=float,
help='confidence threshold')
parser.add_argument('-nt', '--nms_thresh', default=0.5, type=float,
help='NMS threshold')
parser.add_argument('--topk', default=40, type=int,
help='NMS threshold')
parser.add_argument('-K', '--len_clip', default=16, type=int,
help='video clip length.')
parser.add_argument('-m', '--memory', action="store_true", default=False,
help="memory propagate.")
return parser.parse_args()
def multi_hot_vis(args, frame, out_bboxes, orig_w, orig_h, class_names, act_pose=False):
# visualize detection results
for bbox in out_bboxes:
x1, y1, x2, y2 = bbox[:4]
if act_pose:
# only show 14 poses of AVA.
cls_conf = bbox[5:5+14]
else:
# show all actions of AVA.
cls_conf = bbox[5:]
# rescale bbox
x1, x2 = int(x1 * orig_w), int(x2 * orig_w)
y1, y2 = int(y1 * orig_h), int(y2 * orig_h)
# score = obj * cls
det_conf = float(bbox[4])
cls_scores = np.sqrt(det_conf * cls_conf)
indices = np.where(cls_scores > args.vis_thresh)
scores = cls_scores[indices]
indices = list(indices[0])
scores = list(scores)
if len(scores) > 0:
# draw bbox
cv2.rectangle(frame, (x1, y1), (x2, y2), (0, 255, 0), 2)
# draw text
blk = np.zeros(frame.shape, np.uint8)
font = cv2.FONT_HERSHEY_SIMPLEX
coord = []
text = []
text_size = []
for _, cls_ind in enumerate(indices):
text.append("[{:.2f}] ".format(scores[_]) + str(class_names[cls_ind]))
text_size.append(cv2.getTextSize(text[-1], font, fontScale=0.5, thickness=1)[0])
coord.append((x1+3, y1+14+20*_))
cv2.rectangle(blk, (coord[-1][0]-1, coord[-1][1]-12), (coord[-1][0]+text_size[-1][0]+1, coord[-1][1]+text_size[-1][1]-4), (0, 255, 0), cv2.FILLED)
frame = cv2.addWeighted(frame, 1.0, blk, 0.5, 1)
for t in range(len(text)):
cv2.putText(frame, text[t], coord[t], font, 0.5, (0, 0, 0), 1)
return frame
@torch.no_grad()
def detect(args, model, device, transform, class_names, class_colors):
# path to save
save_path = os.path.join(args.save_folder, 'demo', 'videos')
os.makedirs(save_path, exist_ok=True)
# path to video
path_to_video = os.path.join(args.video)
# video
video = cv2.VideoCapture(path_to_video)
fourcc = cv2.VideoWriter_fourcc(*'XVID')
save_size = (960, 720)
save_name = os.path.join(save_path, 'detection.avi')
fps = 20.0
out = cv2.VideoWriter(save_name, fourcc, fps, save_size)
# run
video_clip = []
image_list = []
while(True):
ret, frame = video.read()
if ret:
# to RGB
frame_rgb = frame[..., (2, 1, 0)]
print(frame_rgb.shape)
# to PIL image
frame_pil = Image.fromarray(frame_rgb.astype(np.uint8))
# prepare
if len(video_clip) <= 0:
for _ in range(args.len_clip):
video_clip.append(frame_pil)
video_clip.append(frame_pil)
del video_clip[0]
# orig size
orig_h, orig_w = frame.shape[:2]
# transform
x, _ = transform(video_clip)
# List [T, 3, H, W] -> [3, T, H, W]
x = torch.stack(x, dim=1)
x = x.unsqueeze(0).to(device) # [B, 3, T, H, W], B=1
t0 = time.time()
# inference
outputs = model(x)
print("inference time ", time.time() - t0, "s")
# vis detection results
if args.dataset in ['ava_v2.2']:
batch_bboxes = outputs
# batch size = 1
bboxes = batch_bboxes[0]
# multi hot
frame = multi_hot_vis(
args=args,
frame=frame,
out_bboxes=bboxes,
orig_w=orig_w,
orig_h=orig_h,
class_names=class_names,
act_pose=args.pose
)
elif args.dataset in ['ucf24']:
batch_scores, batch_labels, batch_bboxes = outputs
# batch size = 1
scores = batch_scores[0]
labels = batch_labels[0]
bboxes = batch_bboxes[0]
# rescale
bboxes = rescale_bboxes(bboxes, [orig_w, orig_h])
# one hot
frame = vis_detection(
frame=frame,
scores=scores,
labels=labels,
bboxes=bboxes,
vis_thresh=args.vis_thresh,
class_names=class_names,
class_colors=class_colors
)
# save
frame_resized = cv2.resize(frame, save_size)
out.write(frame_resized)
if args.gif:
gif_resized = cv2.resize(frame, (200, 150))
gif_resized_rgb = gif_resized[..., (2, 1, 0)]
image_list.append(gif_resized_rgb)
if args.show:
# show
cv2.imshow('key-frame detection', frame)
cv2.waitKey(1)
else:
break
video.release()
out.release()
cv2.destroyAllWindows()
# generate GIF
if args.gif:
save_name = os.path.join(save_path, 'detect.gif')
print('generating GIF ...')
imageio.mimsave(save_name, image_list, fps=fps)
print('GIF done: {}'.format(save_name))
if __name__ == '__main__':
np.random.seed(100)
args = parse_args()
# cuda
if args.cuda:
print('use cuda')
device = torch.device("cuda")
else:
device = torch.device("cpu")
# config
d_cfg = build_dataset_config(args)
m_cfg = build_model_config(args)
class_names = d_cfg['label_map']
num_classes = d_cfg['valid_num_classes']
class_colors = [(np.random.randint(255),
np.random.randint(255),
np.random.randint(255)) for _ in range(num_classes)]
# transform
basetransform = BaseTransform(img_size=args.img_size)
# build model
model, _ = build_model(
args=args,
d_cfg=d_cfg,
m_cfg=m_cfg,
device=device,
num_classes=num_classes,
trainable=False
)
# load trained weight
model = load_weight(model=model, path_to_ckpt=args.weight)
# to eval
model = model.to(device).eval()
# run
detect(args=args,
model=model,
device=device,
transform=basetransform,
class_names=class_names,
class_colors=class_colors)