MAC-Multi-scale-Attention-Cascade

This is an official code for "MAC: Mutil-scale Attention Cascade for aerial image segmentation", which is based on MMSegmentation open source toolbox of semantic segmentation. This work achieves 69.06 mIoU on iSAID dataset and 73.37 on ISPRS Vaihingen dataset. Accepted by ICPRAM2024, Oral

Content

1.Requirements
2.Dataset
3.Main code
4.Testing
5.Acknowledge
6.Reference

!!!IMPORTANT!!! Before using, please read and be aware of the SoftwareLicenseAgreement_20230807_v1.pdf

Requirements

We have tested our code with

python=3.10.0 pytorch=1.12.1 CUDA=11.4 CuDNN=8.3.2 mmcv-full=1.6.2

To install mmcv and other related prerequisites, please follow the procedure provided in

MAC-mmsegmentation/docs/en/get_started.md

Dataset

We utlized iSAID dataset for benckmark.

iSAID dataset

Original Input images could be download from DOTA-v1.0 (train/val/test)
Data annotations could be download from iSAID (train/val) The detailed dataset preparation procedure is provided in

MAC-mmsegmentation/docs/en/dataset_prepare.md

If you you have downloaded the dataset, please store it at the following path:

MAC-mmsegmentation/data/iSAID

Main Code

The MAC_head.py contains the main part of MAC model, you could find it at:

MAC-mmsegmentation/mmseg/models/decode_heads/MAC_heads.py

Testing

1. we provide our pretrained model checkpoint mac_latest.pth here:
   Google Drive or Baidu Yun (PSW:nn1t)
   
2. After downloading the checkpoint, please store it at the following path:

MAC-mmsegmentation/work_dirs/mac_isaid/mac_latest.pth

1. The configuration file is stored at the following path:
   

MAC-mmsegmentation/configs/_base_/models/mac_isaid.py and MAC-mmsegmentation/configs/mac/MAC_swin_isaid.py

1. To test the performance of provide checkpoint, please run the following command line:
   

cd MAC-mmsegmentation python tools/test.py configs/mac/MAC_swin_isaid.py work_dirs/mac_isaid/mac_latest.pth --eval mIoU mFscore --show-dir work_dirs/mac_isaid/outs/

1. The visulization resuls will be stored at MAC-mmsegmentation/work_dirs/mac_isaid/outs/

Acknowledgment

• This model is based on the MMSegmentation, thanks to the contributors to the project.
  

Reference

``` @misc{mmseg2020, title={{MMSegmentation}: OpenMMLab Semantic Segmentation Toolbox and Benchmark}, author={MMSegmentation Contributors}, howpublished = {\url{https://github.com/open-mmlab/mmsegmentation}}, year={2020} }