Remove dlib for performance improvements #1
Conversation
WeixuanZ
left a comment
WeixuanZ
left a comment
There was a problem hiding this comment.
Face detection works great, but cropping seems to be a bit dodgy now with more accurate detection
@FChippendale could you have a look
fyi I changed the defaults here: if it can't detect a face it'll set the entire image as a bounding box. We can revert this (with some other fix) if it messes with your assumptions in the cropping stage. |
There was a problem hiding this comment.
We shouldn’t copy secrets into the container - on production the env variables will be set in run config, while locally this will be explicitly passed into docker run command through Makefile variable substitution
Related to 9133f98
Ah I missed that commit; let me rebase this branch on master to include it. What happens when |
DLib and imutils were very slow for high resolution images and required a large build on the initial installation. Fortunately, at this resolution, `cv2.dnn` was ~4x faster, more accurate and already install with opencv. Alternatively, I have provided `FaceDetectorCascade` which uses `c2v.CascadeClassifier`. Although this is less accurate than dlib, I have found that it is much faster at the required resolutions.
- The Makefile runs yack in dev mode, because we want http locally. - Dockerfile only copies required files to the container image. - docker run called with --rm; there's no reason for the container to persist.
antroseco
force-pushed
the
remove_dlib
branch
from
9cf3af2 to
1cb41c0
Compare
antroseco
force-pushed
the
remove_dlib
branch
from
4695fd3 to
aae889c
Compare
In my tests,
cv2.dnnis faster and more accurate than dlib (seeFaceDetectorDNN). I've also addedFaceDetectorCascadefor faster but less accurate face detection.Please check that
requirements.txtwon't break production