When train with a feature map of 512x64x64, the convolutions of h and w are fixed to the dimension of 64.
Then we do the validation with multi-scale which changes the input image, then the spatial size of feature map is not 64x64.
Is it a mistake? or some operations are secretly done? or I have a mistake of understanding the experiments due to I'm a newbie to the task of segmentation?
When train with a feature map of 512x64x64, the convolutions of h and w are fixed to the dimension of 64. Then we do the validation with multi-scale which changes the input image, then the spatial size of feature map is not 64x64.
Is it a mistake? or some operations are secretly done? or I have a mistake of understanding the experiments due to I'm a newbie to the task of segmentation?