Training SegNet on datasets dominated by one category (e.g. Pascal)

[mtsakaguchi]

Hello,

I am trying to train segnet on my own dataset as a preparation before attempting to train segnet on the Pascal VOC dataset. My dataset includes 3 categories, one of them being a "background" category where a large portion (50%-80%) of the image is labelled as "background" (similar to Pascal). I am able to successfully setup the network (using modified train/solver/inference prototxt files based on the ones gone over by the SegNet tutorial) however during after training, the predicted image is largely labelled as "background". Can anyone explain to me what parameters I should be looking into changing (I am somewhat new to caffe)?

Any help would be appreciated.

[shiyuangogogo]

hello, @mtsakaguchi have you solved the problem? I met the same problem as yours. After trained, the model segment almost every target into person (except background). I have add "class_weighting", but it seems useless. could you please give me some advice?

[ghost]

FWIW, trying out a different class_weighing might help (implementing the one in the original segnet paper i.o. my custom one helped me).