Fast-SCNN is a powerful tool for semantic segmentation tasks, popular in machine learning circles for its efficiency and effectiveness. This guide will walk you through the installation, training, evaluation, and demoing of Fast-SCNN using PyTorch.

Table of Contents

• Installation
• Datasets
• Training Fast-SCNN
• Evaluation
• Demo
• Results
• To Do
• References

Installation

Getting started with Fast-SCNN requires a few steps to set up your environment:

• Ensure you have Python 3.x installed. It is recommended to use Anaconda3.
• Install PyTorch 1.0 by selecting your environment on their website and running the appropriate command. For example:

conda install pytorch torchvision cudatoolkit=9.0 -c pytorch

• Clone this repository to your local machine.
• Download the dataset by following the instructions in the next section.
• Note: For training, we currently support Cityscapes and plan to add VOC and ADE20K datasets.

Datasets

You can download the Cityscapes dataset from here. Specifically, download:

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

Training Fast-SCNN

To train Fast-SCNN using the Cityscapes dataset:

python train.py --model fast_scnn --dataset citys

By default, the script looks for the dataset in the “.datasets/citys” directory. You can modify the parameters in the train.py file as needed.

Evaluation

To evaluate your trained model, simply run:

python eval.py

This command provides insights into your model’s performance over the dataset.

Demo

To see the model in action, you can run a demo with a sample image using:

python demo.py --model fast_scnn --input-pic berlin_000000_000019_leftImg8bit.png

Results

Upon completing training, you’ll observe results similar to:

Method	Dataset	crop_size	mIoU	pixAcc
Fast-SCNN (paper)	Cityscapes	768	54.84%	92.37%
Fast-SCNN (ours)	Cityscapes	768	54.84%	92.37%

Note that the results are based on a crop size of 768, which may vary from those reported in the paper.

To Do

• [ ] Add distributed training
• [ ] Support for the VOC, ADE20K datasets
• [ ] Support TensorBoard
• [x] Save the best model
• [x] Add Ohem Loss

References

Rudra PK Poudel. et al. Fast-SCNN: Fast Semantic Segmentation Network.

Troubleshooting

If you encounter any issues during installation or running the code, consider the following troubleshooting tips:

• Ensure all dependencies are correctly installed, especially PyTorch.
• Verify that your dataset is correctly downloaded and in the right directory.
• If using GPU, make sure you have the appropriate CUDA version installed.

For more insights, updates, or to collaborate on AI development projects, stay connected with fxis.ai.

At fxis.ai, we believe that such advancements are crucial for the future of AI, as they enable more comprehensive and effective solutions. Our team is continually exploring new methodologies to push the envelope in artificial intelligence, ensuring that our clients benefit from the latest technological innovations.