Can't be cleaned up, is it an issue with the pre-trained model?

[wxllyf]

FGT

other

What steps have I missed? To shorten the training time, can I train only on the sequence frames of a single video?

[hitachinsk]

To use FGT model, you should provide a mask to indicate the regions you want to remove, otherwise the model cannot know which region should be removed. Moreover, FGT cannot be trained with a single video because the model needs more data to achieve a better result.

[wxllyf]

Both screenshots use the same frames and masks, so it feels strange. I learned about FGT from recommendations in the comments of another project. Below is the original image.

The command I used is:python video_inpainting.py --path images --path_mask masks --outroot test --imgH 512 --imgW 288

[hitachinsk]

It's so strange, because the original content is left in the inpainted frame. If the mask covers all the content in all the frames, the unwanted content should not present in the result. Since FGT propagates the content across all the frames in the videos, please check the masks across all the frames carefully. I think some of the content are not be covered in some of the frames, therefore the result degrades.

[wxllyf]

Reference in

Thank you very much, I will compare and compare what went wrong.