When multi-scale training is applied, the output feature map size also changes, so shouldn't the size of the prior anchor be modified at the same time? But why does anchors remain the same in the code?
it is because of view operation.code is
global_average_pool_reshaped = \
global_average_pool.permute(0, 2, 3, 1).contiguous().view(bsize,
-1, cfg.num_anchors, cfg.num_classes + 5)
in darknet.py
When multi-scale training is applied, the output feature map size also changes, so shouldn't the size of the prior anchor be modified at the same time? But why does anchors remain the same in the code?