-
Notifications
You must be signed in to change notification settings - Fork 0
/
demo.py
229 lines (183 loc) · 6.67 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
import warnings
from os import path
import cv2
import lintel
import numpy as np
import torch
from mmdet.apis import inference_detector, init_detector
from mmdet.structures import DetDataSample
from mmengine.registry import init_default_scope
from datasets.dataset import buildDataset
from funcs.misc import loadConfig, loadCustomized, moveToDevice, setSeed
from mmpose.apis import inference_topdown, init_model
from models.model import buildModel
warnings.filterwarnings("ignore")
detector_config = (
"./mmdetection/configs/faster_rcnn/faster-rcnn_r50-caffe_fpn_ms-1x_coco-person.py"
)
detector_ckpt = (
"./assets/keypoint/faster_rcnn_r50_fpn_1x_coco-person_20201216_175929-d022e227.pth"
)
pose_config = "./mmpose/configs/wholebody_2d_keypoint/topdown_heatmap/coco-wholebody/td-hm_hrnet-w48_dark-8xb32-210e_coco-wholebody-384x288.py"
pose_ckpt = (
"./assets/keypoint/hrnet_w48_coco_wholebody_384x288_dark-f5726563_20200918.pth"
)
part2Index = {
"pose": list(range(11)),
"hand": list(range(91, 133)),
"mouth": list(range(71, 91)),
"face_others": list(range(23, 71)),
}
for key in ["mouth", "face_others", "hand"]:
part2Index[key + "_half"] = part2Index[key][::2]
part2Index[key + "_1_3"] = part2Index[key][::3]
def getSelectedIndexs(videoFrames, numFrames=64):
pad = None
if videoFrames >= numFrames:
start = (videoFrames - numFrames) // 2
selectedIndex = np.arange(start, start + numFrames)
else:
remain = numFrames - videoFrames
selectedIndex = np.arange(0, videoFrames)
padLeft = remain // 2
padRight = remain - padLeft
pad = (padLeft, padRight)
return selectedIndex, pad
def loadFrameNumsTo4DArray(videoBytes, frameNums):
decodedFrames, width, height = lintel.loadvid_frame_nums(
videoBytes, frame_nums=frameNums
)
decodedFrames = np.frombuffer(decodedFrames, dtype=np.uint8)
decodedFrames = np.reshape(decodedFrames, newshape=(-1, height, width, 3))
return decodedFrames
def padArray(videoArrays, pad):
padLeft, padRight = pad
if padLeft > 0:
pad_img = videoArrays[0]
pad = np.tile(
np.expand_dims(pad_img, axis=0),
tuple([padLeft] + [1] * (len(videoArrays.shape) - 1)),
)
videoArrays = np.concatenate([pad, videoArrays], axis=0)
if padRight > 0:
pad_img = videoArrays[-1]
pad = np.tile(
np.expand_dims(pad_img, axis=0),
tuple([padRight] + [1] * (len(videoArrays.shape) - 1)),
)
videoArrays = np.concatenate([videoArrays, pad], axis=0)
return videoArrays
def loadVideo(videoPath):
video = cv2.VideoCapture(videoPath)
videoFrames = 0
while True:
ret, _ = video.read()
if not ret:
break
videoFrames += 1
print(videoFrames)
selectedIndex, pad = getSelectedIndexs(videoFrames)
print(selectedIndex, pad)
with open(videoPath, "rb") as f:
videoBytes = f.read()
videoArrays = loadFrameNumsTo4DArray(videoBytes, selectedIndex) # T,H,W,C
if pad is not None:
videoArrays = padArray(videoArrays, pad)
videoArrays = torch.tensor(videoArrays).float() # T,H,W,C
videoArrays /= 255
videos = []
videos.append(videoArrays)
videos = torch.stack(videos, dim=0).permute(0, 1, 4, 2, 3) # 1,T,C,H,W
return videos
def detectionInference(model, frames):
init_default_scope("mmdet")
detections = []
for frame in frames:
result = inference_detector(model, frame[0])
assert isinstance(result, DetDataSample)
instances = result.pred_instances
instances = instances[instances.scores >= 0.75] # type: ignore
detections.append(instances.bboxes.cpu().numpy()) # type: ignore
return detections
def poseInference(model, frames, detectionResults, filterKeys):
init_default_scope("mmpose")
results = np.zeros((len(frames), 133, 3), dtype=np.float32)
for i, (frame, detectionResult) in enumerate(zip(frames, detectionResults)):
result = inference_topdown(
model, frame[0], bboxes=detectionResult, bbox_format="xyxy"
)
instances = result[0].pred_instances
visibility = instances.keypoints_visible[0][:, np.newaxis] # type: ignore
combined = np.hstack((instances.keypoints[0], visibility)) # type: ignore
combined = combined.reshape((133, 3))
results[i] = combined
filtered = []
for key in sorted(filterKeys):
selected = part2Index[key]
filtered.append(results[:, selected])
filtered = np.concatenate(filtered, axis=1)
filtered = torch.from_numpy(filtered).float() # T,N,3
keypoints = []
keypoints.append(filtered)
keypoints = torch.stack(keypoints, dim=0) # 1,T,N,3
return keypoints
def main():
config = loadConfig()
config["device"] = "cuda:0"
torch.cuda.set_device(config["device"])
setSeed(config["training"]["random_seed"])
dataset = buildDataset(config["data"])
vocab = dataset.vocab
num = len(vocab)
wordEmbTab = []
if dataset.wordEmbTab:
for w in vocab:
wordEmbTab.append(torch.from_numpy(dataset.wordEmbTab[w]))
wordEmbTab = torch.stack(wordEmbTab, dim=0).float().to(config["device"])
del vocab
del dataset
model = buildModel(config, num, wordEmbTab=wordEmbTab)
modelPath = path.join(
"assets",
str(config["data"]["num"]),
"best.ckpt",
)
stateDict = torch.load(modelPath, map_location="cuda")
loadCustomized(model, stateDict["model_state"], verbose=True)
detectorModel = init_detector(
config=detector_config,
checkpoint=detector_ckpt,
device=config["device"],
)
poseModel = init_model(
config=pose_config,
checkpoint=pose_ckpt,
device=config["device"],
)
videoArrays = loadVideo("./videos/learn.mp4")
frames = videoArrays[0].numpy().transpose(0, 2, 3, 1) * 255 # [T,H,W,3]
frames = np.uint8(frames)
assert frames.shape
frames = np.split(frames, frames.shape[0], axis=0)
detectionResults = detectionInference(detectorModel, frames)
poseResults = poseInference(
poseModel, frames, detectionResults, config["data"]["use_keypoints"]
)
st = 64 // 4
end = st + 64 // 2
videos = []
videos.append(videoArrays)
videos.append(videos[-1][:, st:end, ...])
keypoints = []
keypoints.append(poseResults)
keypoints.append(keypoints[-1][:, st:end, ...])
batch = {
"videos": videos,
"keypoints": keypoints,
}
moveToDevice(batch, config["device"])
model.eval()
outputs = model(batch["videos"], batch["keypoints"])
print(outputs)
if __name__ == "__main__":
main()