Make cityscapes binarymask config consistent to polygon config

[krumo]

Hi, I notice that maskrcnn will throw error when running with Cityscapes' binarymask config.

/pytorch/aten/src/THCUNN/ClassNLLCriterion.cu:106: void cunn_ClassNLLCriterion_updateOutput_kernel(Dtype *, Dtype *, Dtype *, long *, Dtype *, int, int, int, int, long) [with Dtype = float, Acctype = float]: block: [0,0,0], thread: [31,0,0] Assertion t >= 0 && t < n_classes failed.

RuntimeError: copy_if failed to synchronize: cudaErrorAssert: device-side assert triggered

However, maskrcnn_benchmark could run with the polygon config succefully. I think the reason is that to reuse the Cityscapes evaluation toolkit, we have to include 'caravan' and 'trailer' in the training classes. Thus, to make maskrcnn could run smoothly with Cityscapes binarymask config, I change the NUM_CLASSES from 9 to 11 and make the experiment setting consistent to the polygon config.

With the changes in the commit, I could train maskrcnn with binarymask config successfully now.