Face Embedder

Purpose

This model outputs a 512-dim vector (embeddings) that represent the human face (facial features).

Architecture

This is a ResNet101 with Arcface loss with margin 0.2 and class weights to deal with class imbalance.

Intended Use

General purpose face embedding model.

Limitations

The model might not work very well on extreme pose / occlusion or blur.

Model Owner

Clarifai Research

Training/Test Data and Taxonomy

The model is trained on face_msceleb1m and finetuned with vggface2_traintest_with_clarifai_lmks_and_5_point_alignment.

Test data is LFW and Janus dataset for evaluation janus-c_identities

Model Evaluation Results

Accuracy on IJBC (1-1 eval): 0.924

Accuracy on LFW: 0.995