ERFNET-PyTorch

A PyTorch implementation for ERFNet, for Semantic Segmentation on Pascal VOC.

Table of Contents:

• ERFNET-PyTorch
  • Project Structure
  • Data Preparation
  • Model
  • Experiment Configs
  • Usage
  • Results
  • Requirements
  • Future Work
  • References
  • License

Project Structure:

├── agents
|  └── erfnet.py # the main training agent 
├── graphs
|  └── models
|  |  └── erfnet.py  # model definition for semantic segmentation
|  |  └── erfnet_imagenet.py  # model definition for imagenet
|  └── losses
|  |  └── loss.py # contains the cross entropy 
├── datasets  # contains all dataloaders for the project
|  └── voc2012.py # dataloader for Pascal Voc dataset
├── data
├── utils # utilities folder containing metrics , config parsing, etc
|  └── assets
├── main.py
├── run.sh

Data Preparation:

We used Pascal Voc 2012 data; to prepare them for training, do the following:

• Visit this, download SBD and PASCAL VOC 2012
• Extract them, you will get benchmark_RELEASE and VOCdevkit folders.
• Download seg11valid.txt and add it into VOCdevkit/VOC2012/ImageSets/Segmentation
• Put the benchmark_RELEASE and VOCdevkit folders in a folder named VOC.
• Remove the color map in ground truth annotations by using remove_gt_colormap.py
• Then Check ./utils/voc_utils.py and ./datasets/voc2012.py for iterating on the VOC Dataset.
• Note that all images has been resized to (256x256) for consistency.

Model:

We are using the same model architecture as given in the paper.

Experiment Configs:

We used the pretrained model on imagenet for ERFNet encoder and trained the model on Pascal VOC for 70 epochs only.

- Input size: 256x256x3
- Batch size: 64
- Learning rate: 5e-4
- learning_rate_patience: 100
- Betas for Adam: 0.5 and 0.999
- Number of epochs: 70

Usage:

• To run the project, you need to add your configurations into the folder configs/. An example of the configurations that should be passed can be found here
• sh run.sh
• To run on a GPU, you need to enable cuda in the config file.

Results:

We only ran a single experiment and got the following results:

Metric	Reproduced
Accuracy	88.75%
Mean-IOU	55.50%

Segmented Images after training:

Requirements:

• Pytorch: 0.4.0
• torchvision: 0.2.1
• tensorboardX: 1.2

Check requirements.txt.

Future Work:

• Train ERFNet with weighted loss

References:

• PyTorch official implementation: https://github.com/Eromera/erfnet_pytorch

License:

This project is licensed under MIT License - see the LICENSE file for details.