about data augmentation.

[czy341181]

您好，非常感谢您的工作。我想问一些关于数据增强方面的问题。

1.我有注意到你在faster_rcnn_RoITrans_r50_fpn_1x_dota1_5_gap512_msTrainTest_rotationTrainTest.py 这个文件中提到了用ms(muti-scale)多尺度的训练方式， 但是在train_config中的img_scale=(1024, 1024)只有一个尺度，请问你是在将数据split的时候，用了多尺度的方式吗？

2.为什么不使用mmdetection中 img_scale=[(1024, 1024),(608, 608),(816,816)] 这种在网络训练过程中随机选取一个尺度的形式呢？

3.在test.py函数中 是直接输入整张图进行预测的，如果test图片比较大，要resize到（1024，1024）会丢失很多信息，为什么在test函数中没有采用demo_large_image.py中将大图crop成小图的形式呢？

期待您的回复！

[dingjiansw101]

1. Yes, we do the multi-scale augmentation during the spliting.
2. The reason why we do not use the img_scale in mmdetection: (1) By using the img_scale, we need to split the images into (1024, 1024) patches first, which may cut some objects into several parts. (2) We apply the resizing operation on the images first, then crop (1024, 1024) patches from it, which can ease the problem of cutting objects into parts.
3. During testing, the large images are cropped into (1024, 1024) patches, so that the resolution will not be lost.