High resolution 高分辨率網絡
中科大和微軟亞洲研究院,發佈了新的人體姿態估計模型,刷新了三項COCO紀錄,還中選了CVPR 2019。
這個名叫HRNet的神經網絡,擁有與衆不同的並聯結構,可以隨時保持高分辨率表徵,不只靠從低分辨率表徵裏,恢復高分辨率表徵。如此一來,姿勢識別的效果明顯提升:
在COCO數據集的關鍵點檢測、姿態估計、多人姿態估計這三項任務裏,HRNet都超越了所有前輩。
https://github.com/icicle4/SuperResulation-HRNet
人體姿態,有權重
https://github.com/stefanopini/simple-HRNet
這個網絡384*288,cpu1.8s
1070 gpu 130ms,特別慢
data=torch.ones(1, 3, 384*3, 288*3).to(device)
分辨率*3,則需要270ms
if __name__ == '__main__':
# model = HRNet(48, 17, 0.1)
model = HRNet(32, 17, 0.1)
# print(model)
# model.load_state_dict(
# # torch.load('./weights/pose_hrnet_w48_384x288.pth')
# torch.load('./weights/pose_hrnet_w32_256x192.pth')
# )
# print('ok!!')
if torch.cuda.is_available():
torch.backends.cudnn.deterministic = True
device = torch.device('cuda:0')
else:
device = torch.device('cpu')
print(device)
model = model.to(device)
data=torch.ones(1, 3, 384, 288).to(device)
import time
for i in range(20):
start=time.time()
y = model(data)
print(y.shape,'time',time.time()-start)
https://github.com/HRNet/HigherHRNet-Human-Pose-Estimation
人臉關鍵點:224 100ms
https://blog.csdn.net/jacke121/article/details/100640305
分割:
https://github.com/HRNet/HRNet-Semantic-Segmentation
https://github.com/HRNet/HRNet-FCOS
能檢測攝像頭
https://github.com/HRNet/HRNet-FCOS/blob/master/demo/webcam.py