多尺度鲁棒性不好，训练数据随机裁剪原图的一部分

[wuzuowuyou]

大家有遇到过吗？用自己的原图测试效果不好，需要把图片resize小尺寸效果才可以
查看源码数据预处理部分是在原图上面随机裁剪512区域进行训练的，不是原图resize到512的，这样做有什么好处呢？

[wuzuowuyou]

左边是原图，右边是裁剪512，把目标车牌都裁减没了，这样合适吗？还有颜色变换也会使得图像面貌全非

[xingyizhou]

Hi,
The crop only happens in training. If you want to turn it off, you can --not_rand_crop --scale 0 --shift 0

[wuzuowuyou]

您好，谢谢您的回复，我还有一个疑问，就是在推理的时候，还是需要经过透视变换trans_input = get_affine_transform(
c, s, 0, [input_w, input_h])， 并且后续就是基于这个透视变换之后的图片的，那得到的坐标是基于透视变换之后的图，那么不需要映射到原图吗？
按我理解，推理的时候直接resize不就可以了吗？

Hello, thank you for your reply. I have another question, that is, when inference, we still need to go through the perspective transformation trans_ input = get_ affine_ transform(
c, s, 0, [input_ w, input_ h] ), and the following is based on the image after the perspective transformation, and the coordinates obtained are based on the perspective transformation image, so do not need to be mapped to the original image?
According to my understanding, when inference, it is OK to directly resize?

[xingyizhou]

Hi,
You are correct. You can use get_affine_transform or directly resize. The reason for me to use get_affine is to follow the setup in CornerNet.