Segmentation network training and inference

This repository contains allows you to train a segmentation network (DeepLabV2) which is based on ResNet-101.

After the network is trained, you can infering pixel-wise masks for given new images and evaluate the accuracy of these masks by using mean Intersection-over-Union (mIoU) metric.

Remember to change all files' paths to your own paths.

Prerequisite

• Python 3.7, PyTorch 1.1.0, and more in requirements.txt
• PASCAL VOC 2012 devkit

Usage

Install python dependencies

pip install -r requirements.txt

Download PASCAL VOC 2012 devkit

• Follow instructions in http://host.robots.ox.ac.uk/pascal/VOC/voc2012/#devkit

train the segmentation network

python train_seg.py

• ALL 3 executing scripts exist in the script directory.
• The segmentation network can be trained by the ground truth masks or the pseudo-masks generated in a weakly supervised case.

Infering masks (on validation set) by the trained segmentation network

python infer.py

• An available pretrained weights (in fully supervised case, i.e., trained by ground truth masks) is online: https://drive.google.com/file/d/12oLbmwwEbD8vpGH7QwogSggHNS4eTSiT/view?usp=sharing (mIoU on validation set is 77.78%).

Evaluation the infered masks using mIoU metric

python results_evaluation.py

Acknowledgement

This repository is based on work from: https://github.com/kazuto1011/deeplab-pytorch. Thanks for their impressive work.