Reproducing the output of https://github.com/BU-Spark/ml-atfal-mafkoda-missing-children pipeline with the following considerations in mind:

1. Use REST API in order to be able to offload work on multiple servers/worker nodes
2. Use CPU wherever possible to reduce cost

In order to do so, some networks will be replaced with lighter networks. OpenVino and onnx-runtime will be used wherever an implementation that provides similar quality is found.

1. /faces: Endpoint that will return boxes for detected faces within an image url
2. /enhance: Given an image URL and a box that denotes a face, crop and enhance (super resolution) the face box
3. /frontalize: Given an enhanced face from step #2, frontalize the face and return a frontalized image

More endpoints in progress..

The API can run in a docker container. The following steps can be used to build the image.

Build

sudo docker build -f docker/Dockerfile-x64 . -tatfal-ai

And run it..

sudo docker run --publish 8080:8080 -it atfal-ai:latest

You can now open your browser to: http://127.0.0.1:8080/docs and experiment with the APIs

To run on aarch64, you can use the Dockerfile-aarch64

sudo docker build -f docker/Dockerfile-aarch64 . -tatfal-ai

Refer to the wiki for deployment on Oracle Cloud Free Tier

• Consider CFPGAN for face restoration and alignment https://github.com/TencentARC/GFPGAN
• Port cfrgan to onnx or OpenVino

1. Wherever a third-party library is included due to the need to patch it for CPU, the code within the third-party library will follow the license model for that library.
2. Whevever a third-party code file is included, the code for that file will follow the original license model for that file. An example of this is the OpenVino demo files used in the pipeline for face detector and face enhancer.