search

vt-vl-lab / FGVC

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

test question #65

Open gao123qiang opened 2 years ago

gao123qiang commented 2 years ago

  1. I observed that when the scene changes greatly, the repair results are better, but when the scene changes little, there maybe have artifacts in the repair results. How to deal with them? Can you give some suggestions

  2. I don't know how these command parameters affect the results?

gaochen315 commented 2 years ago
  1. When you say "the scene changes greatly", do you mean the camera motion is large, or the objects in the scene have a large motion?

Camera motion: if the camera is barely moved, we do not have a lot to propagate. The completion quality mainly depends on single image inpainting, and the result could be bad. object motion: If the object moves too fast, RAFT may fail due to rapid motion. And the propagation quality will be degraded.

  1. If you want to remove objects or inpaint the missing region, use --mode object_removal. Use --mode video_extrapolation for extrapolation. --seamless will give you seamless results but it requires a longer processing time.
gao123qiang commented 2 years ago

@gaochen315 what i mean is the Camera motion, the result likes you said, i notice the single image inpaint is too old, do you plan to update it with v2 or other ? now i try to add a new method: CTSDG to inpaint image , it will be over, 1, can you give me some advises where i should pay attention to ? 2, can you share me about the args: edge_guide,Nonlocal, alternate_corr, mixed_precision and consistencyThres, how to affect the results ? i have tested it with different style, but i found less change in result.