search

wshilton / andrew

GNU General Public License v3.0
0 stars 0 forks source link

Face discrimination scheme #4

Open wshilton opened 1 year ago

wshilton commented 1 year ago

Determine if there exists a suitable open source model for purposes of human facial discrimination in the sense of a comprehensive scheme for storage, retrieval, normalizations, class biases, etc. If not, devise a basic such scheme using linear synthetic discriminants or a similar procedure amenable to the nearly real-time requirement.

wshilton commented 1 year ago

There are a plethora of open source models. M-RCNN appears to be suitable for facial attribute estimation. The work here is then to integrate M-RCNN facilities with a human identity-level discriminator.

wshilton commented 1 year ago

DeepFace is sufficient for recognition and verification. Its facilities for FAE, which use a pre-trained 3-layer convolutional network for reduction to a single categorical metric with only seven possible states, are insufficient for our purposes, however. For this reason, DeepFace will be adopted for recognition and verification and M-RCNN or a suitable alternative will perform the higher fidelity FAE.

wshilton commented 1 year ago

For FAE, candidate starting point is Google's MediaPipe face landmark detection.

wshilton commented 1 year ago

DeepFace facilities for classification minimize surface-level norms. Since this was the primary value-add, we are now considering alternatives.

wshilton commented 1 year ago

DeepFace facilities for classification minimize surface-level norms. Since this was the primary value-add, we are now considering alternatives.

Some quick work could be done that reformulates the classical linear synthetic discriminants as mapping from a contour of the face landmarks. Still, a storage and retrieval scheme would need to formulated or identified -- this problem class probably already has standardized models. Design work and literature review now follows.