Training on the MVSEC dataset about dceiflow

[danqu130]

Thank you for your interest in our work. I hope I can solve your concerns.

1. The training data for the model trained with MVSEC are all valid indexs of outdoor_day2, i.e., [1950, 28500].
2. The experiments in our paper are training 100 epochs on FlyingChairs2 and 200 epochs on MVSEC. For MVSEC, batch_size is 16, and the other hyperparameters remain unchanged.
3. Since a longer time has passed, I will try to find it.

I apologize for the unclear expressions in the paper. I think I need to clarify some additional experimental details.

The experiments on the MVSEC dataset only were additionally requested by the reviewers and were not the focus of our paper. Therefore we made a fair comparison with baselines only and not with existing methods in Table 4. In the experiments on MVSEC, we not only used the full outdoor_day2 data, but also add image data augmentation as well as an unsupervised training loss function (cf. ARFlow), in order to achieve acceptable generalization performances. The weights of supervised and unsupervised losses are 1:2.

[Sisyphus-ZPJ]

Thanks for your reply. It is very helpful to me!