arbitrary interpolation training

[HamidGadirov]

I have a question about arbitrary-timestep frame interpolation. As explained in the paper, this was achieved by the model RIFEm by directly inputting t to the IFNet (which creates a mask of the t values, as I understood from the code). For the data, you randomly select 3 frames from Vimeo90K-Septuplet and calculate the t, as explained in the paper. But are there any other changes in the training stage? Did the model converge immediately? I am experimenting with arbitrary interpolation with a method that uses RIFE as a backbone and it is difficult to get proper convergence on my datasets in case of arbitrary interpolation training.

[hzwer]

Hello, I have some experience:

1. In my previous experiments, I first tried the selection method of (0, x, 6) and found that the effect was not good; random selection of (x, y, z) (x < y < z) worked much better
2. Another change is to appropriately reduce the coefficient of optical flow loss (I'm not sure if you have something similar).
3. I found that the l1 loss will drop significantly slower than the original setting, but in the end it is indeed possible to use a model to interpolate different t.