faceMaskDetection-SSD-FasterRcnn-MosaicAug

7

Dataset

Models

Deep Learning FrameWork

  • Pytorch

Implementation Details

  • The Main motivation in developing Facemask Detection is to deploy model in mobile Device and Heroku therefor i need to reduce the detection latency low as possible while keeping decent MAP. Therefor i have given more emphasis on SSD Model with MobileNetV3 large as backbone.
  • I trained pytorch pretraind ssd_mobilenet_v3_large object detection model with heavy augmentation, given 0.38 MAP. The Model was trained for above 100 epochs.
  • The pretrained model separately trained with Mosaic Augmentation but i have trained for 40 EPOCHS and got 0.39MAP. Model performance can be further improved by for more epochs.

Deployment

  • Model is deployed to Heroku using Flask.
  • Using REST API i have developed an android application , where user can capture picture from camera and can detect mask.
  • Another android application is developed where The Pytorch scripted model is deployed on device.
  • Check out the URLS below for related repos.

Other Repo URLS