By looking into the align process of deepface, we found it using simple method so that sometimes it cannot handle complex face detection & alignment (E.g One-side face detection & alignment), thus we tried to use the alignment from InsightFace and apply it into DeepFace. It only requries minor change of DeepFace source code, and we need this repo to maintain the alignment and the source code.
The represent() function in DeepFace does not take alignment as parameter, thus the modified DeepFace.py was proposed in this repository. You can use your own faces for testing by setting the alignment argument to False in DeepFace and importing modified DeepFace.py.
pip install -r requirements.txt
import face_align
face_aligned = face_align.get_face_align(img_file)
import DeepFace
#Turning off align in deepface, also remember to turn off detector.
#analyze face
arcanalyze = DeepFace.analyze(face_aligned, actions = ['age', 'gender', 'race', 'emotion'],detector_backend='skip',align=False)
#face embedding(512*1 dim)
arcrep = DeepFace.represent(face_aligned, model_name = 'ArcFace', model = None, enforce_detection = True, detector_backend = 'skip', align = False, normalization = 'base')
React
Typescript
Django
Laravel
Facebook
Microsoft
Google
Alibaba
D3
Tencent