Minimum device hardware to run face detection using WebGL as a backend on tensorflowjs

irufano · January 26, 2024, 6:50am

I want to use tensorflowjs face detection pretrained model into browser using WebGL as a backend, is there any minimum device requirement like vga, ram etc, to run this on browser?

Jason · January 29, 2024, 7:53pm

I am unaware of the lower bounds of execution here. The model itself is pretty small - I think just a few MB if I remember correctly, so as long as you have enough GPU memory for that then it should work. The only question is what speed - which depends on the hardware running on - however assuming you dont run out of memory or such then it should still run at some FPS. I have personally tested this model on devices as old as Pixel 5 and it runs fast enough for real time applications for me to give you some context.

Please note we have 2 implementations of FaceMesh model:

github.com

tensorflow/tfjs-models/blob/master/face-landmarks-detection/README.md

# Face Landmarks Detection

This package provides models for running real-time face detection and landmark tracking.

Currently, we provide 1 model option:

#### MediaPipe:
[Demo](https://storage.googleapis.com/tfjs-models/demos/face-landmarks-detection/index.html?model=mediapipe_face_mesh)

MediaPipe Facemesh can detect multiple faces, each face contains 468 keypoints.

More background information about the package, as well as its performance characteristics on different datasets, can be found here: [Model Card](https://drive.google.com/file/d/1QvwWNfFoweGVjsXF3DXzcrCnz-mx-Lha/preview). The facemesh package optionally loads an iris detection model, whose model card can be found here: [Model Card](https://drive.google.com/file/d/1tV7EJb3XgMS7FwOErTgLU1ZocYyNmwlf/preview).

-------------------------------------------------------------------------------
## Table of Contents
1. [How to Run It](#how-to-run-it)
2. [Keypoint Diagram](#keypoint-diagram)
3. [Example Code and Demos](#example-code-and-demos)

-------------------------------------------------------------------------------

This file has been truncated. show original

MediaPipe version

TensorFlow.js version

You may prefer to use one over the other based on your needs. I would recommend using the Mediapipe JS example as of 2024 as we no longer support dev on the TFJS version of the model - this was made in a time when there was a need for that due to GPU compatibility issues back then on iOS etc.

irufano · February 27, 2024, 1:56am

thanks for your response, btw can we retrain the pretrained models like face-detection short range models?