🚘 The easiest implementation of fully convolutional networks

• Task: semantic segmentation, it's a very important task for automated driving

• The model is based on CVPR '15 best paper honorable mentioned Fully Convolutional Networks for Semantic Segmentation

Results

Trials

Sub-models performance

Performance

Trained with two popular benchmark dataset: CamVid and Cityscapes

dataset	n_class	pixel accuracy
Cityscapes	20	96%
CamVid	32	93%

Training

Install packages

pip3 install -r requirements.txt

and download pytorch 0.2.0 from pytorch.org

and download CamVid dataset (recommended) or Cityscapes dataset

Run the code

• default dataset is CamVid

create a directory named "CamVid", and put data into it, then run python codes:

python3 python/CamVid_utils.py 
python3 python/train.py CamVid

• or train with CityScapes

create a directory named "CityScapes", and put data into it, then run python codes:

python3 python/CityScapes_utils.py 
python3 python/train.py CityScapes

Developed for Deep learning course / Denmark's Technical University / Fall 2019

Students

Dimitrios Dagdilelis / @dimidagd

Mitul Vekariya

Inspired by

Po-Chih Huang / @pochih

Known issues

• Imbalanced classes in the dataset, (small objects have less pixels), need to balance the loss function, possible solution or here

Useful urls

plotting