PSENet
PSENet copied to clipboard
whai362
Is resnet 50 used as backbone in ICPR MTWI 2018 Challenge 2? or other
in ICPR MTWI 2018 Challenge 2, your result is F = 75.2， which do you select as backbone, resnet50,resnet101 or resnet152?
the backbone is resnet 152，i think without ic17 pretrain model is also ok，because icpr has enough images.
i use fpn as your done, but the result is bad. as far as you know, how much can FPN improve?
you can use pspnet or other common segmentation framework. the fpn is not necessary.
all we need is a segmentation result，kernel and the algorithm of expanding.
I mean the key idea is the kernel mechanism and expanding method，the segmentation framework is alternative.
all we need is a segmentation result，kernel and the algorithm of expanding.
your paper said: The backbone of PSENet is implemented from FPN, F = C(P 2 , P 3 , P 4 , P 5 ) = P 2 k Up ×2 (P 3 ) k Up ×4 (P 4 ) k Up ×8 (P 5 ). i wonder whether my result with fpn is reasonable ? is FPN helpful to improve detection result?
is FPN helpful to improve detection result？Yes， fpn is help to make accurate segmentation. But there are some details in the code. So if you are trying to build a psenet-like method easily，i think implement from other segmentation framework (e.g. pspnet) is also ok. But pspnet need more gpu memory，and the speed is slower.
