Insightface Facial Recognition
JackyWong88 opened this issue · comments
Previously, you've mentioned that Insightface uses 512-D vectors of facial data here: #22 . However, when retrieving the vector data, it only shows 128 floats of data. Is this the expected result?
For facial recognition, the similarity percentage for Insightface can be rather low when comparing some Angelina pictures with other celebrities (I pulled pictures from here: https://www.kairos.com/demos). When running facial similarities among all the celebrities from the prior link using your code for Insightface, the similarities between 2 of the same person has a max of about 46% similarity and a min of 24%, which are both rather low. I was wondering if during your tests, you have noticed an optimal threshold for recognizing that a person is the same person and eliminating any false positives.
Sources of all models: https://github.com/php-opencv/php-opencv-examples/tree/master/models
So I got this model from there https://github.com/axinc-ai/ailia-models/tree/master/face_identification/insightface
Because I was needed to have onnx-version. If you have onnx-version of insightface with 512-D vectors you can use it instead 128. Or you can ask about optimal threshold in that project https://github.com/axinc-ai/ailia-models