Fast-SCNN
A PyTorch implementation of Fast-SCNN: Fast Semantic Segmentation Network from the paper by Rudra PK Poudel, Stephan Liwicki.
Installation
- Python 3.x. Recommended using Anaconda3
- PyTorch 1.0. Install PyTorch by selecting your environment on the website and running the appropriate command. Such as:
conda install pytorch torchvision cudatoolkit=9.0 -c pytorch - Clone this repository.
- Download the dataset by following the instructions below.
- Note: For training, we currently support cityscapes, and aim to add VOC and ADE20K.
Datasets
- You can download cityscapes from here. Note: please download leftImg8bit_trainvaltest.zip(11GB) and gtFine_trainvaltest(241MB).
Training-Fast-SCNN
- By default, we assume you have downloaded the cityscapes dataset in the
./datasets/citysdir. - To train Fast-SCNN using the train script the parameters listed in
train.pyas a flag or manually change them.
python train.py --model fast_scnn --dataset citys
Evaluation
To evaluate a trained network:
python eval.py
Demo
Running a demo:
python demo.py --model fast_scnn --input-pic './png/berlin_000000_000019_leftImg8bit.png'
Results
| Method | Dataset | crop_size | mIoU | pixAcc |
|---|---|---|---|---|
| Fast-SCNN(paper) | cityscapes | |||
| Fast-SCNN(ours) | cityscapes | 768 | 54.84% | 92.37% |
Note: The result based on crop_size=768, which is different with paper.
TODO
- [ ] add distributed training
- [ ] Support for the VOC, ADE20K dataset
- [ ] Support TensorBoard
- [x] save the best model
- [x] add Ohem Loss
