vt-vl-lab / FGVC

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

extrapolation artifacts #31

Closed ak9250 closed 3 years ago

ak9250 commented 3 years ago

in google colab running extrapolation on the first example frame tennis has artifacts 49e6bdb8-d461-466c-b131-5ee1ea41a651

ak9250 commented 3 years ago

@gaochen315 also artifacts are present with or without --seamless

ak9250 commented 3 years ago

@gaochen315 any update on this?

gaochen315 commented 3 years ago

This is expected. For example, the top right region is never revealed in the video. So we have to count on the single image completion. The current image completion is trained for inpainting, so it will perform poorly on image outpainting. Hope this helps.

ak9250 commented 3 years ago

@gaochen315 ok thanks