How to input muti-channel label image,such as(width*height*channel) and How to use label_class_selection.prototxt

[TITC]

Dear author,

1.questions 1

I try to input multi-channel as input label but I can't find any description about it. And I have tried a lot of way to input muti-channel image

such as :

• use Caffe default layer slice to segment one label to two
• use HDF5 layer to substitute for PathData layer
• use multi-top-label as input label(I have tweaked some code which correlated with hdf5write )

but all of them are failed

---

2.questions 2

I know that semantic segmentation task's label is binary image, but when I saw the prototxt file

cdeep3m/model/inception_residual_train_prediction_1fm/label_class_selection.prototxt

Line 1 in abf9958

ignore_rest_of_label:true

this file is for label mapping, so does this meaning cdeep3m support multilabel image as input and can bound different pixel value and label number for mapping? but I don't know how to use it, and can't find any description of it?

Could you please provide some hints to me? Thanks for your help.

Yuhang Tao
Best wishes to you.