CornerNet-Lite学习篇（上）

普林斯顿大学的团队出的新论文
github已开源，指路CornerNet-Lite github
本篇是实现记录：

依赖环境

• Python 3.7
• PyTorch 1.0.0
• CUDA 10
• GCC 4.9.2 or above

配置环境按照github上很好实现
测试下时间

本地环境
GTX1080

~/CornerNet-Lite$ python demo.py 

demo.py 加了个时间输出

#!/usr/bin/env python

import cv2
import time
from core.detectors import CornerNet_Squeeze
from core.vis_utils import draw_bboxes

detector = CornerNet_Squeeze()
image    = cv2.imread("demo.jpg")

for i in range(100):
   start_time = time.time()
   bboxes = detector(image)
   end_time = time.time()
   print ("used: " + str(end_time - start_time))
image  = draw_bboxes(image, bboxes)
cv2.imwrite("demo_out.jpg", image)

结果:

used: 0.04870462417602539
used: 0.05021524429321289
used: 0.05280447006225586
used: 0.049675941467285156
used: 0.04538917541503906
used: 0.08332228660583496
used: 0.052958011627197266
used: 0.10060548782348633
used: 0.0784444808959961
used: 0.06858110427856445
used: 0.05750775337219238
used: 0.05352616310119629
used: 0.08527612686157227
used: 0.08002758026123047
used: 0.050453901290893555
used: 0.09716010093688965
used: 0.06107616424560547
used: 0.04721355438232422
used: 0.04533505439758301
used: 0.04773664474487305
used: 0.044405221939086914
used: 0.0460507869720459
used: 0.0438234806060791
used: 0.04536104202270508
used: 0.04605841636657715
used: 0.04491996765136719

所以Squeeze的运行时间大概是40-50ms
out: