Generate edge-TPU compiled TF-lite from Face Mask detector model for Coral

[mhejrati]

[mrn-mln]

@mhejrati I generated tflite model and sent a PR for this task on neuralet-models repo.

[mhejrati]

@mrn-mln Cool, thanks.

Can you please provide more details of what you have done there and what are the next steps?

[mrn-mln]

@mhejrati A ssd-mobilenet-v1 detector is trained on face-mask dataset. I converted the model to quantized int8 model which is compatible with edgetpu devices (the mAP = 0.67%). At the next step, I am going to use the tflite model for implementing a face/face-mask detector app at the social-distancing repo. The task is related to issue 106.

[mrn-mln]

I tested a ssd-movilenet-v1, the generalization of the model was not well. So, I will try to train a new model for this task. I will update my activity here.

[mrn-mln]

I trained resnet-50 on WIDER FACE dataset. The mAP was ~0.52 which was not well enough. I will try to train a model on the HollywoodHead detection dataset in order to detect head.