What is the output modified using one-hot code if I want to train model by 3D images? Such as video frames

[che006]

I stacked three frames as a training sample. The semantic labels of every frames were included in my dataset. But I dont know how to modify the output layer with one-hot code. The 2D labels is different from 3D.