DNN face detector fails with images that have many faces

[Maik Qualmann]

Yes, the problem that not all faces are always detected has been reported and 
we also have bug reports. I think Thanh did his work with the YOLO model. The 
YOLO model recognizes all faces in pictures. But because of the size of the 
precalculated model data, we decided on the SSDMOBILENET. Therefore we would 
like to make the YOLO model available for download and implement it as part of 
the GSoC2020.

You can easily switch to the YOLO model (facedetector.cpp: 85). The model data 
must be installed to /usr/share/digikam/facesengine/.

Maik

Am Samstag, 16. Mai 2020, 23:24:39 CEST schrieb Nghia Duong:
> Hi all,
> 
> I was messing around with the face detection module when I found some
> abnormal behaviors. DNN face detection module fails to detect faces in some
> images that contain many faces, especially images that have faces align
> horizontally. Furthermore, the overloaded functions prepareForDetection of
> OpenCVDNNFaceDetector give different performances. For example, It cannot
> detect any faces in the sample photo codingsprint2009-01.jpg in the folder
> core/tests/facesengine/data.
> 
> Q: Has this kind of issue ever been reported? If not, could any of you help
> me re-verify this problem?
> 
> Thank you in advance.
> 
> Best regards.
> Nghia