majiayao-roborock / slam-paper-reading

0 stars 0 forks source link

NetVLAD #2

Open majiayao-roborock opened 4 years ago

majiayao-roborock commented 4 years ago

原理

Screenshot from 2020-02-17 17-27-06 可以看到针对不同的相似度的图片结果可以表现相似程度,在随后的测试中,现有的依赖街景得到的模型,可以在我们常见的办公区,室外环境使用,阈值建议在0.9~1,当两张图片的结果差<阈值时,则可以认为来到了同一地点。NetVLAD输入是彩色图像,模型大小在1G左右,480×960 1050ti 运行时间在 0.67s。 MobileNetVLAD输入是灰色图像,模型大小在60M左右,480×960 1050ti 运行时间在 0.027s。

ref NetVLAD与MobileNetVLAD 极简的打开方式 https://zhuanlan.zhihu.com/p/76475388