Final Project for DS4440 - Practical Neural Networks at Northeastern University (Fall 2020).

Reimplementation of DeepLabv3+ with a modified ResNet-50 backbone as specified in DeepLabv3.

The main results can be viewed in evaluate.ipynb.

Folder structure and base classes were generated from pytorch-template by Victor Huang.

• ResNet
• DeepLabv3
• DeepLabv3+
• Cityscapes Dataset Paper

• Cityscapes scripts

Download the following files from Cityscapes: - gtFine_trainvaltest.zip - leftImg8bit_trainvaltest.zip - gtCoarse.zip - leftImg8bit_trainextra.zip

Extract the files into ./data/ (or wherever you specify as the data_dir in config.json). Make sure to use the fine annotations for the train and val sets and coarse annotations for the train_extra set:

data/
│
├── gtFine/
│   ├── train/
│   └── val/
│
├── gtCoarse/
│   └── train_extra/
│  
└── leftImg8bit/
    ├── train/
    ├── train_extra/
    ├── val/
    └── test/

pip install -r requirements.txt

# TODO

# image
python detect.py

# video
python detect_video.py

python train.py

# track performance
tensorboard --logdir=experiments/runs/{exper_name}/{run_id}

python test.py

• Track more metrics (Dice Score and iIoU)
• Investigate other loss functions (RMI)