Multi-Scale-Dilation Net is a deep neural network designed for semantic segmentation task, which integrates multi-scale feature extraction by the dilated convolution and the reduced FCN8s backbone. It serves as one of the baseline methods for the UAVid dataset.
First, clone the repository and go to the MSDNet folder.
git clone https://github.com/YeLyuUT/MSDNet.git
cd MSDNet
mkdir data
ln -s <path to UAVid dataset> data/uavid
bash init.sh
python toolkit/getCroppedTrainData.py
Our model is tested with early tensorflow1.3 version and python2.7
You can follow the instruction on the official site to download.
Link: https://www.tensorflow.org/install/pip?lang=python2
python LMain_msd.py -m t -f fileListTrainCropped.txt
python LMain_msd.py -m t -f fileListTrainCropped.txt -w <checkpoints>
python LMain_msd.py -m tp -f fileListTrainCropped.txt -w <checkpoints>
python LMain_msd.py -m p -f fileListTest.txt -w <checkpoint weight>
We offer the pretrained weights in the google drive link below, which should give 57.0% mIoU score on the UAVid2020 benchmark.
https://drive.google.com/drive/folders/1TzWeAu9Kb3Eqh3Y2YdyZ-3g4_zJinhvA?usp=sharing
Simply download the checkpoint files to the 'checkpoints' subfolder, and run the following command,
python LMain_msd.py -m p -f fileListTest.txt -w checkpoints/model.ckpt
Please cite our paper if you find the work useful.
@article{LYU2020108,
author = "Ye Lyu and George Vosselman and Gui-Song Xia and Alper Yilmaz and Michael Ying Yang",
title = "UAVid: A semantic segmentation dataset for UAV imagery",
journal = "ISPRS Journal of Photogrammetry and Remote Sensing",
volume = "165",
pages = "108 - 119",
year = "2020",
issn = "0924-2716",
doi = "https://doi.org/10.1016/j.isprsjprs.2020.05.009",
url = "http://www.sciencedirect.com/science/article/pii/S0924271620301295"}