The pipeline of the algorithm is inspired by [27
]. However, due to the difference between reverberation artifacts and shadow artifacts, we use probabilistic networks and a novel transform function in our work. The pipeline can be divided into three parts: (1) probabilistic segmentation on hard labels, (2) transform function,and (3) probabilistic segmentation on soft labels. We use the hard labels to train the hard label segmentation network, then transform the output of this network to soft labels by the transform function. After that, we use the soft labels to train the soft label segmentation network.