search

cheind / dest

:panda_face: One Millisecond Deformable Shape Tracking Library (DEST)
BSD 3-Clause "New" or "Revised" License
288 stars 109 forks source link

DEST vs DLIB #14

Open sahil-ag opened 6 years ago

sahil-ag commented 6 years ago

Quick question, will both DEST and DLIB give same output? Also, which will be more accurate.

cheind commented 6 years ago

Good question which I cannot answer thoroughly. However, both DLIB and DEST have been applied to the same video.

DEST https://www.youtube.com/watch?v=Hewjc0oyqPQ DLIB https://www.youtube.com/watch?v=oWviqXI9xmI

I don't know how DLIB video was made, but for DEST the following procedure was applied. Viola-Jones algorithm was used to determine a bounding rectangle containing a face. Regression was then applied to this rectangle. Since Viola-Jones takes around 50ms per frame, DEST uses the regressed frame of frame n-1 and uses that rectangle to initialize the bounding box for frame n.