search

xiong-zhitong / pytorch_RFCN

279 stars 54 forks source link

the code problem #2

Closed NHZlX closed 6 years ago

NHZlX commented 7 years ago

I didn't run your code, but I have read the source code and found that self.rfcn_score = Conv2d(1024,7*7*8, 1,1, bn=False) self.rfcn_bbox = Conv2d(1024, 7*7*self.n_classes,1,1,bn=False)

Shouldn't it be

self.rfcn_score = Conv2d(1024,7*7*self.n_class, 1,1, bn=False) self.rfcn_bbox = Conv2d(1024, 7*7*8,1,1,bn=False)

?

DuinoDu commented 7 years ago

I agree with you, @NHZlX

CosmosShadow commented 6 years ago

This help me. Training on another dataset, this will works wrong. Thank you.

yangwf1 commented 6 years ago

why should it be self.rfcn_bbox = Conv2d(1024, 778,1,1,bn=False), and not Conv2d(1024, 774,1,1,bn=False) as clarified in the paper

princewang1994 commented 6 years ago

@yangwf1 in this implement, first 4 channels(0 to 3) are set for "background bbox prediction" even it is no used in training (see variable rpn_bbox_inside_weights , it is to mask bg background boxes), because there is no boxes target for bg, it seems useless in bbox training. And following 4-7 channels are for fg boxes regression.

btw, My R-FCN.pytorch has released for R-FCN and CoupleNet implement, welcome staring :)