search

yijingru / BBAVectors-Oriented-Object-Detection

[WACV2021] Oriented Object Detection in Aerial Images with Box Boundary-Aware Vectors
MIT License
469 stars 88 forks source link

你好，我训练hrsc后，测试的可视化里检测框都是平行四边形，而不是像标签那样的旋转矩形，请问是怎么回事呢 #41

Closed 18804601171 closed 4 years ago

yijingru commented 4 years ago

It's a result of affine transformation when the output aspect ratio is different from the training images.

18804601171 commented 4 years ago

@yijingru 那怎么能让检测结果是正常的旋转矩形呢，因为我训练完dota后，测试的时候，我看检测的可视化里检测框基本都是正常的旋转矩形呢

yijingru commented 4 years ago

Because dota is from cropped images, whose aspect ratio is 1. You may pad zero to the input images to keep the aspect ratio and make the transformation of the rectangle to be rigid. But in this way, you may sacrifice either computation or resolution.

seacloud-0420 commented 3 years ago

@18804601171我也遇到了这个问题，我用自己的数据集没有裁剪成600*600，检测结果也是平行四边形，请问你解决了吗？ @yijingru 不太明白你回答的，可以详细说下吗？

yijingru commented 3 years ago

图片宽高比的问题，1:1就是矩形

seacloud-0420 commented 3 years ago

那我应该怎么做才能是矩形呢？

------------------ 原始邮件 ------------------ 发件人: "yijingru/BBAVectors-Oriented-Object-Detection" @.>; 发送时间: 2021年8月21日(星期六) 下午2:37 @.>; @.**@.>; 主题: Re: [yijingru/BBAVectors-Oriented-Object-Detection] 你好，我训练hrsc后，测试的可视化里检测框都是平行四边形，而不是像标签那样的旋转矩形，请问是怎么回事呢 (#41)

图片宽高比的问题，1:1就是矩形

— You are receiving this because you commented. Reply to this email directly, view it on GitHub, or unsubscribe. Triage notifications on the go with GitHub Mobile for iOS or Android.

yijingru commented 3 years ago
  1. 保持training和testing图片宽高比一致，使用padding?
  2. 或者修改decoding box方式？强行转换矩形，比如找平行边垂直线，同时穿过对边的检测点。如果有此类算法，欢迎分享！