vt-vl-lab / FGVC

[ECCV 2020] Flow-edge Guided Video Completion
Other
1.55k stars 263 forks source link

Terrible FOV Extrapolation results #43

Closed Korbpw closed 3 years ago

Korbpw commented 3 years ago

https://user-images.githubusercontent.com/46784998/114278144-5f067880-99fc-11eb-8392-6778379037bc.mp4

gaochen315 commented 3 years ago

This result is expected for this demo sequence because few frames are provided.

Our algorithm first propagates content across frames. The result will look plausible if the content is available for propagation. For example, the left region of the first frames and the right region of the last frames looks good. For those never-seen regions, we have to apply an off-the-shelf inpainting algorithm to fill in the content. Please note that this algorithm is designed for inpainting, not outpainting. Thus the extrapolation results will suffer. The terrible results you observed are caused by this inpainting algorithm.

I will improve the extrapolation performance if I find a good outpainting algorithm. Please stay tuned. Thank you!