Constant probability for object detection

[xternalz]

Hi,

Thanks for this nice work. It seems that you are using a constant GridMask probability for object detection. The set_prob function only gets called when cfg.GRID.PROB == 0. Does that work better than linearly increasing the probability?

Thanks.

[akuxcw]

Hi, in our experiments, we didn't observe obvious promotion when using linearly schedule of the probability in detection. We think it may because the backbone already uses pre-trained weights in detection.

[xternalz]

Thanks.

[hedes1992]

Hi, in our experiments, we didn't observe obvious promotion when using linearly schedule of the probability in detection. We think it may because the backbone already uses pre-trained weights in detection.

Hello, have you do experiments using official pretrained model from pytorch, but not the pretrained model with GridMask from imagenet? If only use official pretrained model, I guess the linearly schedule is more better? Thank you

[akuxcw]

Hello, We always use the official pretrained model from pytorch, which means we use the same pretrained model for baseline and our results. We didn't use pretrained model with GridMask from imagenet. Because we want to show it is gridmask that can improve the result of detection instead of the pretrained model.

[hedes1992]

Hello, We always use the official pretrained model from pytorch, which means we use the same pretrained model for baseline and our results. We didn't use pretrained model with GridMask from imagenet. Because we want to show it is gridmask that can improve the result of detection instead of the pretrained model.

Thanks for your reply. Ok, so you mean we can use prob=0.7 directly in detectino training because the official model is good enough?(In table7 of the grid mask paper).

[akuxcw]

Yes.