SwiftNet

Source code to reproduce results from

Efficient semantic segmentation with pyramidal fusion
Marin Oršić, Siniša Šegvić
Pattern Recognition, 2020.

In Defense of Pre-trained ImageNet Architectures for Real-time Semantic Segmentation of Road-driving Images
Marin Oršić*, Ivan Krešo*, Siniša Šegvić, Petra Bevandić (* denotes equal contribution)
CVPR, 2019.

Steps to reproduce

Install requirements

• Python 3.7+

pip install -r requirements.txt

Download Cityscapes

From https://www.cityscapes-dataset.com/downloads/ download:

• leftImg8bit_trainvaltest.zip (11GB)
• gtFine_trainvaltest.zip (241MB)

Either download and extract to datasets/ or create a symbolic link datasets/Cityscapes

Expected dataset structure for Cityscapes is:

gtFine/
    train/
        aachen/
            aachen_000000_000019.png
            ...
        ...
    val/
        ...
leftImg8bit/
    train/
        aachen/
            aachen_000000_000019.png
            ...
        ...
    val/
        ...

Evaluate

Pre-trained Cityscapes models available

• Download and extract to weights directory.

Set evaluating = True inside config file (eg. configs/rn18_single_scale.py) and run:

python eval.py configs/rn18_single_scale.py

Train

python train.py configs/rn18_single_scale.py --store_dir=/path/to/store/experiments

---

Additional instructions for students of D307:

• Use config configs/rn18_single_scale.py for your experiments
• If you need to use another dataset, create a symbolic link to the folder inside dataset/custom_dataset and add support for the dataset in data/custom_data/custom_data.py (use Cityscapes files as template)
• You do not have to install and set up Cython since we have added new methods for confusion matrix and metrics calculation
• Be careful when naming the experiments folders so you do not override something. Output folder should contain saved best model and last model, pickle file with validation mIoU and loss values.
• When evaluating, be sure to set evaluating = True in config and provide correct path of the saved model.