Page 1 of 1

Poor quality of the alignment file after Extract

Posted: Wed Dec 11, 2019 12:09 pm
by uequalsn
Hi All,

I wonder if there's a way to streamline the extraction process in terms of the quality of the output alignment file. Tried to extract the face from a video in mkv format using recommended Detector and Aligner (S3Fd and Fan, respectively). Ran extraction with two different maskers (VGG-Obstructed, Extended). The results are roughly the same: I literally have to click through all the frames and fix displaced face contours.

While I can imagine that it may be hard for the algo to detect a face in a dynamic scene, but this is not my case. Any suggestions are appreciated!

Re: Poor quality of the alignment file after Extract

Posted: Sat Dec 14, 2019 5:05 am
by bryanlyon
Alignments do not have to be exact, especially when using a smart mask. In those situations, the alignments are just used to get the face generally upright.

Re: Poor quality of the alignment file after Extract

Posted: Sat Dec 14, 2019 1:30 pm
by djandg
Ive seen that sometimes the detector picks up false faces and aligns, so two or maybe more alignments can be on a frame. Especially of its a crowded video. So I normaly end up running manual alignment to get ride of these "ghosts". Time consuming but thats the way it is, A coffee machine is a great "ghost" :lol:

Re: Poor quality of the alignment file after Extract

Posted: Sun Dec 15, 2019 8:15 pm
by bryanlyon
Yes, and our guide covers both the alignment tool and the sort tool. The sort tool is definitely going to be the faster and easier option for doing this.