Can yolo heads deal with input images where width <> height?
Created by: flishwang
Currently, yolo_box only support input shape [N, C, H, W] where H == W. However, This means that I have to pad or expand the images (i.e. 512*1024) to square, resulting in either less compute efficiency or extra inaccuracy. So can you add support for inputs with different shape?
In chinese: 目前，fluid.layers中yolo_box的输入尺寸只允许图片宽==高。因此，对于长宽比并不接近的图片来说，使用padding补成正方形的话，计算效率会受影响，使用拉伸方式拉成正方形的话，因为形变过大，会影响结果精度。能否增加对宽高不一致的输入尺寸的支持呢？