Liveness Detection: Analyzing Faces at a Distance

135 views
Skip to first unread message

Darrell Bellert

unread,
Jun 6, 2022, 10:32:16 AM6/6/22
to doubango-ai
I am interested in detecting liveness for faces that are NOT close to the camera, which I realize is in direct violation of the First Commandment at https://www.doubango.org/SDKs/face-liveness/docs/Improving_recall.html.

Our environment is constrained such that the user cannot get close to the camera.  Can the liveness detection model still be of any value in such an environment?  I've tweaked some of the json config settings to no avail, such as min face size thresholds, but I'm not entirely sure where these are all documented and what they all mean.

Attached is a sample image from our environment.
62993f48266c060e29df9492.jpeg

Mamadou DIOP

unread,
Jun 6, 2022, 11:04:24 AM6/6/22
to Darrell Bellert, doubango-ai
One processing done on the faces is texture analysis. When the face is very far and the image quality very low it become impossible to perform texture analysis which means the face will likely be declared as spoof.
To improve the correct detection of far away faces you’ll need to improve the image quality.

Another processing done is checking some distances about your eyes, nose, ears… In the attached image we don’t clear see your eyes. If the same face was taken very close the eyes would be easier to check. So, make sure to look at the camera if you want a correct detection, otherwise we’ll think it’s a wax figure and this is why the attached image is detected as “disguise”.

Unfortunately we haven’t added the documentation about the configuration entries but we’ll do it. Change “ detect_face_minsize” and “ detect_face_minsize”:

We’re working on v1.0 and the results we have are much much better than the current version (v0.6). That version works better on far away and low resolution faces. It should be release this week or in the worse scenario the next one.

This said, I don’t know any company that can accurately do liveness detection on far away  faces. Even if you are using hardware (3D sensor, depth…) the distance should be reasonable.

We receive a lot of data from our customers which help us improve the product. If you have sample data and willing to share it with us, you’re welcome.

--
You received this message because you are subscribed to the Google Groups "doubango-ai" group.
To unsubscribe from this group and stop receiving emails from it, send an email to doubango-ai...@googlegroups.com.
To view this discussion on the web visit https://groups.google.com/d/msgid/doubango-ai/e8f34c5b-d406-4331-9247-36a5f8980689n%40googlegroups.com.
<62993f48266c060e29df9492.jpeg>

Reply all
Reply to author
Forward
0 new messages