Skip to content
Switch branches/tags

Latest commit


Git stats


Failed to load latest commit information.
Latest commit message
Commit time


This repository contains the implementation details of our paper: [arXiv:1809.06323]

"Efficient Dense Modules of Asymmetric Convolution for Real-Time Semantic Segmentation"

by Shao-Yuan Lo, Hsueh-Ming Hang, Sheng-Wei Chan, Jing-Jhih Lin


NOTE: I cannot share our source codes due to the sponsor's request for this project. If you are interested in reproducing EDANet, you can try to write the network code (we use PyTorch) based on the layer disposal I provided here. Or you can access the implementations by others: [pytorch_EDANet] (Pengfei Wang).

If you have any questions or suggestions, welcome to contact Shao-Yuan Lo (me, the email address is shown in the paper).


This repository is organized as:

  • models contains the layer disposal of EDANet and each network variant introduced in our paper.
  • train contains the training details and the class weights used to train our networks.

We replace our source codes with lists to show the detailed network architectures and training setup. In our actual implementation, PyTorch (v0.4.1) is used to perform all the experiments.

Performance on the Cityscapes dataset

EDANet achieves a mIoU of 67.32 on the Cityscapes test set without any pretrained model. (Cityscapes leaderboard)

For an input resolution of 512x1024, EDANet can run at the speed of 108 FPS on a single GTX 1080Ti GPU and 81 FPS on a single Titan X GPU.


In our paper, it also shows the evaluation results on the CamVid dataset.


If EDANet is useful for your research, please consider citing our paper:

  • S.-Y. Lo, H.-M. Hang, S.-W. Chan, and J.-J. Lin, “Efficient dense modules of asymmetric convolution for real-time semantic segmentation,” in ACM International Conference on Multimedia in Asia (MMAsia), 2019. [Best Paper Award]


No releases published


No packages published