can no longer detect far faces and somtimes close faces of any person in the scene #83

Open
opened 2021-08-04 12:55:33 +02:00 by ozmix · 6 comments
ozmix commented 2021-08-04 12:55:33 +02:00 (Migrated from github.com)

I have been trying this out with different videos and images, sometimes it detects the faces and sometimes it does not. before the improved mask update, it uses to recognize the face on all videos and images. now it sometimes cuts out either being too far away and too close, it's very strange. I'm not sure what it could be, however, I do like this project your doing and I believe it can be the best of the best.

I have been trying this out with different videos and images, sometimes it detects the faces and sometimes it does not. before the improved mask update, it uses to recognize the face on all videos and images. now it sometimes cuts out either being too far away and too close, it's very strange. I'm not sure what it could be, however, I do like this project your doing and I believe it can be the best of the best.
NNNNAI commented 2021-08-06 09:06:07 +02:00 (Migrated from github.com)

This should be caused by the mask-dectected model could not detect the mask of that face which is too far or too close. You can try to remove the --use_mask augument in the demo code to see what is different.

This should be caused by the mask-dectected model could not detect the mask of that face which is too far or too close. You can try to remove the --use_mask augument in the demo code to see what is different.
woctezuma commented 2021-08-06 09:45:06 +02:00 (Migrated from github.com)

I guess this happens if the mask-based model does not find the face here:
1b8f482795/util/reverse2original.py (L107)

It makes sense that the recall gets worse with an additional filtering step.

I guess this happens if the mask-based model does not find the face here: https://github.com/neuralchen/SimSwap/blob/1b8f48279570dda55aa0843987d9a7b6ee87af81/util/reverse2original.py#L107 It makes sense that the recall gets worse with an additional filtering step.
ozmix commented 2021-08-06 13:21:18 +02:00 (Migrated from github.com)

gonna try it now see what happens

gonna try it now see what happens
ozmix commented 2021-08-06 13:34:52 +02:00 (Migrated from github.com)

gonna try it now see what happens

so I tried it and it has definitely worked to a degree but im gonna try it some more on other videos and images

> gonna try it now see what happens so I tried it and it has definitely worked to a degree but im gonna try it some more on other videos and images
lschaupp commented 2022-01-04 11:44:19 +01:00 (Migrated from github.com)

I have played around with the current version (of this date posting it) and there are still issues with close up faces. I also reverted the changes back to the date (SimSwap/util/reverse2original.py) before masking was implemented and there were also issues with it. Does anybody know what the issue could be?

I have played around with the current version (of this date posting it) and there are still issues with close up faces. I also reverted the changes back to the date (SimSwap/util/reverse2original.py) before masking was implemented and there were also issues with it. Does anybody know what the issue could be?
woctezuma commented 2022-01-04 12:13:57 +01:00 (Migrated from github.com)

I would suggest creating a new Github issue, and showing some examples of the issues with close-up faces.

As you mentioned, the issue may have been always there, in which case there is not much to do, except improve face detection maybe.

I would suggest creating a new Github issue, and showing some examples of the issues with close-up faces. As you mentioned, the issue may have been always there, in which case there is not much to do, except improve face detection maybe.
Sign in to join this conversation.