Hi, I am confused about the asymmetric motion field figure above, which you put in the paper. Since we only have t=0 and t=1, it is hard to judge whether the motion is symmetric or not. In my opinion, if the right foot in the above image is lower a little bit, then the motion will be symmetric and we don't have to take asymmetric motion field into account.

Hi @JunHeum, Thanks for the wonderful work! I'm trying to re-implement ABME, but I don not know how to implement census loss. Even if in the reference paper I cannot get enough information about it. Could you please kindly provide an implementation of census loss? Thanks a lot!

bicorr6 = self.bilateral_corr6(c16, c26, torch.zeros(c16.size(0), 2, c16.size(2), c16.size(3)).to(c16)) to bicorr6 = self.bilateral_corr6(c16, c26, torch.zeros(c16.size(0), 2, c16.size(2), c16.size(3)).to(c16.device)) ?

I want to reproduce the 0.22s for 480*640 image reported by your paper, could you please provide some code? I implemented some code by myself but failed, about 2 times slower.

We checked SNU-FILM dataset is not available in this link.

Therefore, we share the SNU-FILM dataset until that link provides the dataset available.

You should download two zip files. RGB image files Sequence index files

Hi, thanks for your awesome work.

I have a question as an extension to Issue #8 . I understand that the provided pretrained model can be used recursively to produce results for more moments than t=0.5. However as far as I understood, according to the paper and your responses, it seems like the conclusion leads to: 'although the provided model is for predicting an intermediate frame t=0.5, it can be rebuilt for estimating frame at random t.'

I thus tried to reproduce your model for arbitrary time t based on your code and figured that it doesn't seem possible to be extended for arbitrary time. Specifically, I doubt that Eq. 15 of the paper cannot hold for arbitrary time t. The pixel intensity seems to be intended to sum to 1. The equation holds as long as t=0.5, but for example, when t=0.3, M_t1=0, M_t0=1, the pixel intensity becomes 1.4 . I wanted to ask if there is anything that I misunderstood or got wrong... Please, correct me if I am wrong.

Why use vimeo90k 'testlist.txt' provided for validation. As far as i know, when training the test dataset is not visible.

And another question, what optimizer and lr_scheduler do you use when training? Can give some details? Thanks.

Hi, As what mentioned in the paper, the reliability map Z is computed by where I think, Z is related to the difference about the interped frame in time t by V_(t->0) and V_(t->1). However, in the code,

where frame2_1 is compute by the flow SBM_ * (-1)， which I think is flow from 0 to 1, not time t to 1.
In '''frame2_1, Mask2_1 = warp(frame1_, SBM_ * (-1), return_mask=True)''', frame2_1 will be a warpped frame from frame1 to frame3, not frame2. Hope I have maken a clear statement of my question.

Hi team,

I think the paper mentions it is possible to interpolate at any time t. But the codes provided only perform a t=0.5 interpolation. Is it possible to modify the code and interpolate at a given time t? Thank you!