Xin Tao

The training process include random process, e.g. random data sequence order, random initialization etc. If you want deterministic process, you need to fix random seed for python random module, numpy...

Yes. I think multi-gpu is possible. We use only common modules, which should be easy to put on multi-gpu. But the current codes need to be modified to use multi-gpu.

Hi, We have just uploaded the testing dataset, which consists of 30 sequences of 31 frames. [download](https://tinyurl.com/y7prfsgd)

Because TensorFlow uses static graph, and our network modules needs pre-assigned input size in released models. 1. You may split images into several overlapped regions and test separately, then stitch...

Please at least include more details. Did you test on your own images with different sizes? If you change input sizes, I think you need to modify code accordingly.

Sorry for this late update. We will update training code very soon. As for training data, it cost 42GB, we do not have a good way to share them. If...

It is a network, that takes 2 frames as input, and outputs estimated optical flow. The structure is the same as the optical flow module in the paper 'Real-Time Video...

I implement warp operator based on the code here: https://github.com/tensorflow/models/blob/master/research/transformer/spatial_transformer.py#L145 The difference is I modify parameter theta to flow field. For the second question, I am a little confused. For...

Sure, I will upload those models soon.

We calculate Flow(i->0) and use forward warping to warp i-frame to 0-frame. So that all frames are aligned to 0-frame. The reason is according to imaging model. Please refer to...