search

yousongzhu / CoupleNet

Code for training 'CoupleNet: Coupling Global Structure with Local Parts for Object Detection' in ICCV 2017
61 stars 22 forks source link

Table4 - Global FCN #9

Closed foralliance closed 5 years ago

foralliance commented 5 years ago

@tshizys HI

In Table4，for Global FCN, it can be regarded as a Faster R-CNN. But the depth of roI-wise subnetwork is only two, why did it work out so well( 78.5 )?? In our common sense, roi-wise subnetwork should be very deep to be correct.

yousongzhu commented 5 years ago

@foralliance ,Hi,

Actually, using two layers for roi-wise subnetwork is enough to encode the features in our experiments (also in FPN and Mask R-CNN). The way of using the whole conv5 in ResNet as the roi-wise subnetwork is just to follow the early structure of Faster R-CNN.

foralliance commented 5 years ago

@tshizys thanks!!! get it