RandomCrop in the train pipeline #23
-
Hi, I was reading training config files for several models, and then I saw the following lines. It seems like you guys cropped an image of 480x960 then resized it to 1333x800 with the keep_ratio argument set to True. Then, doesn't it shrink the image because image width(=960) is smaller than target width(=800)? If my understanding is correct, are there specific reasons for padding so many pixels on height? My current understanding came from the following RandomCrop implementation by mmdet. Or am I missing something? |
Beta Was this translation helpful? Give feedback.
Replies: 1 comment 2 replies
-
@junghyun-avikus you misunderstood the |
Beta Was this translation helpful? Give feedback.
@junghyun-avikus you misunderstood the
crop_size
parameter. It does not crop 480x960 images. I copy this line from the link you provided:"absolute_range" uniformly samples crop_h in range [crop_size[0], min(h, crop_size[1])] and crop_w in range [crop_size[0], min(w, crop_size[1])].