3D-mobilestereonet can not reach the test results of kitti2015，why？thanks

[flyxuexi]

I used the MSNet3D_SF.ckpt you provided.We fine-tuned 400 rounds on kitti2015 and submitted the parallax map to the kitti website, but the results were very different from those in the paper. In the paper, it is 1.75 3.87 2.10 1.61 3.50 1.92, but what I measured is 2.01 4.40 2.40 1.85 3.88 2.18, is it the wrong way for me to fine-tune, or is it for some other reason? Please help me.