No input available.


Workflow Graph

face detection → crop → face landmarks → face alignment → face clusterer and face-embedder

Landmarks & Alignment

See more model card details of face-landmarks

five-point face landmarks: left_eye_center, right_eye_center, nose_tip, mouth_left, mouth_right


There are many reasons for degraded face model & workflow performance. Below is a non-exhaustive list

  • Camera type (e.g. IR, black/white, fish-eye/360), lighting/illumintation, and motion blur from videos
  • Image resolution and face size: face sizes that are less than 40px or <10px between pupils might not be detected
  • Face orientation (pan, roll and tilt): In general faces with more than 45 degrees of rotation may not be detected
  • Occlusion and blur: heavy occlusion is going to hurt performance, but the model has been trained with partial occlusion

Open Set Problem

While our general face embedding model is trained with angular margin loss and Adacos loss to generalize better on open set, for custom face embedding model it is strongly recommended to add explicit negatives to turn the problem into a "close set" and softmax loss is ok - as negative mining is more critical in this case.

  • Workflow ID
  • Description
  • Last Updated
    Oct 03, 2022
  • Privacy
  • Share