My implementation of softmax interpolation "Softmax Splatting for Video Frame Interpolation".
My env: Pytorch 1.3, One GPU RTX2080Ti
Train this model SoftSplatModel_v3 on Vimeo90K 3-frames dataset.
Train logs are saved in ./logs
Model weights are saved in ./ckpt
Evaluation on Vimeo90K teset set: 35.92dB
Difference with original paper:
- I compute the warped images using forward warp flow + backwarp image, refer to Line 158 in SoftSplatModel_v3.py
- I use L1 loss for training