search

YuwenXiong / py-R-FCN

R-FCN with joint training and python support
MIT License
1.05k stars 472 forks source link

peformance on smaller network #78

Open newwhitecheng opened 7 years ago

newwhitecheng commented 7 years ago

Hi, I try to train on a smaller network such as CaffeNet and GoogleNet. But the result is map= 0.002, something like that. Do you ever get a chance to train R-FCN with smaller network? If yes, How is the perforamce and can you share your prototxt?

Thank you!

315386775 commented 7 years ago

@newwhitecheng can you share your prototxt?

ananddb90 commented 7 years ago

@newwhitecheng I am also facing the same problem. I used pre-trained ResNet18 and MobileNet but in both the cases I am getting mAP not more than 40% @YuwenXiong Does anyone know if I reduce score map size (from 7 to 3) and number of channel for conv_new_1 (from 1024 to 512), rpn_conv/3x3 (from 512 to 256) OR number of channels are crucial for detection networks

Thank you

newwhitecheng commented 7 years ago

@ananddb90 I don't know the trick of training R-FCN, that's why I initiated this thread. I hope I can help you. But I have no idea.