Training existing data sets

[lin-fangzhou]

It seems that few existing datasets can be used to train patchmatchnet. Even those datasets used for evaluation (such as ETH3D) do not provide enough information. I wonder if anyone uses other existing datasets to train this network.