Home

Awesome

EDANet

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

EDANet

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).

Pakages

This repository is organized as:

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.

sample_results

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

Citation

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