These are 4 models we trained and valiated on our MegaDepth dataset, as described in our updated paper. If you want to compare your algorithms with ours both numerically and qualitative in your paper, please consider using these models.

The reason we trained several models on our dataset is that we found if we do several cross-dataset tests without even validating on any other target dataset, the variance of generalization performance is large aross different dataset. To avoid basis of cross-data set, we decide to use several models we trained on our MegaDepth and compute the average of error. 
