trellixvulnteam/regseg_0jhu

The official implementation of "Rethink Dilated Convolution for Real-time Semantic Segmentation"

RegSeg

The official implementation of "Rethink Dilated Convolution for Real-time Semantic Segmentation"

Paper: arxiv

D block

Decoder

Setup

Install the dependencies in requirements.txt by using pip and virtualenv.

Download Cityscapes

go to https://www.cityscapes-dataset.com, create an account, and download gtFine_trainvaltest.zip and leftImg8bit_trainvaltest.zip. Unzip both of them and put them in a directory called cityscapes_dataset. The cityscapes_dataset directory should be inside the RegSeg directory. If you put the dataset somewhere else, you can set the config field

config["dataset_dir"]="the location of your dataset"

You can delete the test images to save some space if you don't want to submit to the competition. Make sure that you have downloaded the required python packages and run

CITYSCAPES_DATASET=cityscapes_dataset csCreateTrainIdLabelImgs

There are 19 classes.

Results from paper

To see the ablation studies results from the paper, go here.

Usage

To visualize your model, go to show.py.

To see the model definitions and do some speed tests, go to model.py.

To train, validate, benchmark, and save the results of your model, go to train.py.

Cityscapes test results

RegSeg (exp48_decoder26, 30FPS)

test mIOU: 78.3

model weights

Larger RegSeg (exp53_decoder29, 20 FPS)

test mIOU: 79.5

model weights

Comparison against DDRNet-23

Run	RegSeg model weights	DDRNet-23 model weights
run1	77.76	77.84
run2	78.85	78.07
run3	78.07	77.53

Citation

If you find our work helpful, please consider citing our paper.

@article{gao2021rethink,
  title={Rethink Dilated Convolution for Real-time Semantic Segmentation},
  author={Gao, Roland},
  journal={arXiv preprint arXiv:2111.09957},
  year={2021}
}