InsightFace for TFJS

TFJS port of InsightFace

Models

Repository contains pretrained TFJS graph models for the following InsightFace variations

• human-faceres: included for reference, 6.7MB weights
• human-mobilefacenet: included for reference, 5.0MB weights
• insightface-mobilenet-emore: 6.7MB weights, 1.6ms avg
• insightface-mobilenet-swish: 12MB weights, 3.0ms avg
• insightface-ghostnet-strides1: 7.8MB weighs, 9.3ms avg
• insightface-ghostnet-strides2: 7.7MB weights, 7.4ms avg
• insightface-efficientnet-b0: 13MB weights, 9.8ms avg

Notes

• Models have been quantized to F16 for size
• All models take [1, 112, 112, 3] cropped and normalized [0..1] image of a face as input
  and produce as single float array as output which represents face embedding
• Performance numbers are using RTX3060

Demo & Compare

Demo app in /src uses Human library to detect and crop faces from input images before running InsightFace models to calculate face descriptors (embeddings)

And yes, you can use any other face detection method

Sample Images

GitHub repository is void of sample images, beforing running demo place any number of images of any resolution containing one or more faces into /assets/samples/

Sample Screenshot

Run

Use built-in dev server to compile sources and start a web server:

npm run dev

Credits

• Original implementation: https://github.com/deepinsight/insightface
• Keras port: https://github.com/leondgarse/Keras_insightface

TBD

• Optimize similarity in (%) from raw distance
• Find best face.scale per model