High quality implementation of the portrait mode effect using Neural Networks.

Traditionally, the portrait mode effect has been achieved using 2 lenses which detect objects present in the foreground and in the background. With advances in the field of ML, this effect can also be implemented using only image segmentation. Using the pretrained DeepLab-v3+ open source TensorFlow model, we can find the objects in the foreground of the image and blur the background to replicate this effect.

Download DeepLab model trained on COCO dataset from the Model Zoo and convert it to onnx model.
For example download deeplabv3_mnv2_pascal_train_aug_2018_01_29 and convert it to onnx model from frozen_inference_graph.pb

python -m tf2onnx.convert 
  --opset 11
  --fold_const
  --graphdef frozen_inference_graph.pb
  --output deeplabv3_mnv2_pascal_train_aug.onnx
  --inputs ImageTensor:0
  --outputs SemanticPredictions:0

or use already-made onnx model.

Build C# source code, run application and use drag & drop to produce "portrait mode" effect.

Figure 1. Example of "portrait mode" effect

[1] Portrait mode implementation using TF and Python.
[2] Tutorial: how to convert them to ONNX and run them under onnxruntime.
[3] Microsoft: ONNX Runtime C# API.
[4] UMapx.NET - digital signal processing library.
[5] Real-time chromakey effect.