Pytorch Implementation of "Fully Convolutional Network", "Recurrent Residual Convolutional Neural Network based on U-Net (R2U-Net)" and "DeepLabV3" on PascalVOC and Cityscapes dataset.
Navami Kairanda Priyanka Mohanta
Following packages are used
- python 3.8
- pytorch 1.7
- torchvision 0.8.1
- pytorch-lightning 1.2.3
For tasks 2 and 3,
Download and unzip gtFine_trainvaltest.zip (241MB) and leftImg8bit_trainvaltest.zip (11GB) from cityscapes site https://www.cityscapes-dataset.com/downloads/
Generate trainId labels for the dataset, using the scripts provided by Cityscape authors https://github.com/mcordts/cityscapesScripts
git clone https://github.com/mcordts/cityscapesScripts.git
pip install cityscapesScripts
CITYSCAPES_DATASET_PATH=/HPS/Navami/work/code/nnti/R2U-Net/cityscapes/
export CITYSCAPES_DATASET=$CITYSCAPES_DATASET_PATH
python /HPS/Navami/work/code/nnti/cityscapesScripts/cityscapesscripts/preparation/createTrainIdLabelImgs.py
Download resnet pretraineed model from https://download.pytorch.org/models/resnet50-19c8e357.pth and update corresponding path in resnet.py
For task 1, run Vision_task_1.ipynb jupyter notebook
For tasks 2 and 3,
python main.py /path/to/expt/logdir
For tasks 2 and 3, download model from Microsoft Teams
python eval.py /path/to/expt/logdir {model_name}.tar
Task 1: Jonathan Long, Evan Shelhamer, and Trevor Darrell. Fully Convolutional Networks for Semantic Segmentation. arXiv e-prints, page arXiv:1411.4038, November 2014.
Task 2: Md Zahangir Alom, Mahmudul Hasan, Chris Yakopcic, Tarek M Taha, and Vijayan K Asari. Recurrent residual convolutional neural network based on u-net (r2u-net) for medical image segmentation. arXiv preprint arXiv:1802.06955, 2018.
Task 3: Liang-Chieh Chen, George Papandreou, Florian Schroff, and Hartwig Adam. Rethinking atrous convolution for semantic image segmentation. arXiv preprint arXiv:1706.05587, 2017.