Skip to main content

Face ID INT8

A three-stage on-device face recognition pipeline: face detection, 68-point facial keypoints, and 256-D face embeddings.

Supported processors

  • SL2619

Performance

Torq Compiler/Runtime v2.3.0

14.701ms
Inference Time
37.6MB
DRAM Footprint

Face ID

Model Overview​

A three-stage face recognition pipeline optimized for on-device execution with the Synaptics Torq runtime and NPU on Astra® SL2610-Series processors.

The pipeline detects faces, extracts facial landmarks, and generates normalized face embeddings for recognition and enrollment.

Model Features​

  • Model Type: Face detection, facial keypoint detection, and face embedding
  • Quantization: INT8

Face Detection​

  • File: face_detection.vmfb
  • Input: 1x704x1280x1 int8 grayscale
  • Outputs: 1x1100x2 int8 box coordinates (two tensors) and 1x1100x1 int8 confidence scores

Face Keypoints​

  • File: face_keypoint_static.vmfb
  • Input: 1x56x56x3 int8 RGB face ROI
  • Output: 68 facial landmarks and an MAE value

Face Embeddings​

  • File: face_embeddings_static.vmfb
  • Input: 1x112x112x3 int8 aligned RGB face
  • Output: 1x256 int8 face embedding

Model Pipeline​

Input Image
|
Face Detection
|
Face ROI and 68 Keypoints
|
Five-Point Face Alignment
|
256-D Face Embedding

Deployment​

The compiled model files are available for download on Huggingface at Synaptics/face-id-torq.

Usage Tutorial​

A usage example is provided at Torq Examples / Face_ID. Check out the README for instructions.

License​

These models are licensed under AGPL-3.0.

Learn More​