Image Segmentation With Deeplab

Image Segmentation using Deeplab v3+

Summary

Experimenting with Quantization of Tensorflow Models on various datasets with the DeepLab v3 Decoder architecture and MobileNet v2 Encoder architecture using using a variety of techniques including

• Quantization aware training
• Quantization aware training with delay
• Post training Quantization
• Quantized Inference/Evalulation

Presentation Slide

• Lighter, Faster Semantic Segmentation by Post-Training Quantization and Quantization-Aware Training

DeepLab: Deep Labelling for Semantic Image Segmentation

@inproceedings{deeplabv3plus2018, title={Encoder-Decoder with Atrous Separable Convolution for Semantic Image Segmentation}, author={Liang-Chieh Chen and Yukun Zhu and George Papandreou and Florian Schroff and Hartwig Adam}, booktitle={ECCV}, year={2018} }

Installation

pip install all the following required packages.

Requirement

• TensorFlow 1.15
• Jupyter Notebook
• Python 3.6
• Numpy
• Pillow
• matplotlib
• conda

Note: For a ready to use envirenment, a deeplearning ami on an EC2 instance would come with all the required packages needed to run this repo immediatly.

Usage on Colab

• Fine-tuning and Quantization

• Inference

Usage on AWS

• clone the repo
• navigate to ImageSegmentationWithDeeplab (command: cd ImageSegmentationWithDeeplab)
• run the command "jupyter notebook"
• use the provided url (default: localhost:8888)
• open the "inference_deeplab.ipynb" notebook
• From drop down list Cell > Run All

Results

• FLOAT32 Segmentation

• Post-Quantization UINT8 Segmentation (no fine-tuning)

• Post-Quantization UINT8 Segmentation (10K-iteration fine-tuning)

• Quantization-Aware-Training UNIT8 Segmentation