about fine-tuned PWC-net

[andersonhusky]

Thanks for your excellent work again! I found that you guys fine-tune the PWC-net in Sintel and KITTI training datasets, and I found a large performance gap between your fine-tuned PWC-net and the original one in dataset "demo-kitti", could you please share the fine-tuned pareameter file? This will be a huge help for me, thanks!

[vkorotkine]

The fine-tuning of PWC-net is done in the original PWC paper itself as far as I understand.
Did you manage to resolve your issue? What was the cause?

From the original PWC-Net paper, https://arxiv.org/pdf/1709.02371.pdf
"We first train the models using the FlyingChairs dataset
in Caffe [28] using the Slong learning rate schedule introduced in [24], i.e., starting from 0.0001 and reducing the
learning rate by half at 0.4M, 0.6M, 0.8M, and 1M iterations. The data augmentation scheme is the same as that
in [24]. We crop 448 × 384 patches during data augmentation and use a batch size of 8. We then fine-tune the models on the FlyingThings3D dataset using the Sf ine schedule [24] while excluding image pairs with extreme motion
(magnitude larger than 1000 pixels). The cropped image
size is 768 × 384 and the batch size is 4. Finally, we finetune the models using the Sintel and KITTI training set and
will explain the details below"