Compare different neural network architecture for semantic segmentation problem on mobiles.

Android application written in Kotlin and using tensorflow-mobile and tensorflow-lite.
With this application you can load image and segmentate it with choosen Deep Neural Network or run benchmarks for compare average working time.

There are four networks, in this table you can find information about quality and size:

And here you can find some time measurements:

You can find notebooks for each model with pretrained and trained weights, freezed graphs and tflite files. If you want, to run this notebooks, you should install packages from requirements, tensorflow, keras and this repo. You can skip installation of frameworks and use my docker.

It's important to understand, that this project hasn't completed yet. So feel free to ask quiestions in issues.

• Implement E-Net (depends on Tensorflow ScatterNd layer issue);
• Port IC-Net and U-Net to tflite;
• More networks??
• Train on MS COCO;
• ...