Embedded Deep Learning based Semantic Segmentation using TensorFLow Lite

Objective

To create a vision based environment surveying application using deep learning capable of running no embedded devices as well.

Workflow

Training Process

Computer Vision Task : Semantic Segmentation
Dataset : CamVid
Architecture : Endoder-Decoder
Models : UNet, SegNet
Encoder : MobileNetV2

Compression and Inference based on TensorFlow Lite

Compression Techniques : [x] Quantization      [ ] Prunning
Compression Tools : [x] TensorFlow Lite      [ ] TensorRT

Compression Result

Inference using TensorFlow Lite Interpreter