HRL-Hierarchical Relation Learning for Few-Shot Semantic Segmentation in Remote Sensing Images

Option 1: training from scratch Download the pre-trained backbones from resnet50_v2.pth+vgg16_bn.pth

train_base.sh

Option 2: loading the trained models

Put the provided models in the /initmodel

https://github.com/XinnHe/HRL/releases/tag/initmodel%2FBaseNet%2Fisaid https://github.com/XinnHe/HRL/releases/tag/initmodel%2FBaseNet%2Fdlrsd https://github.com/XinnHe/HRL/releases/tag/initmodel%2FBaseNet%2FLoveDA

python train.py --shot 1 --split 0 --dataset iSAID --backbone vgg

python test.py --shot 1 --split 0 --dataset iSAID --backbone vgg

During testing, please manually set the path of the weights via the resume_path argument, for example:
resume_path ="./weights/best_dlrsd_res50_split0_1shot.pth"

• 5 random seeds are used.
• For each seed, 1000 support-query samples are randomly selected.
• test.py is run 5 independent times, and the **average resul Not
• This setting follows DMNet

• 5 random seeds are used.
• For each seed, 1000 support-query samples are randomly selected.
• test.py is run 10 independent times, and the average result is reported in the paper.
• *This setting follows R2Net

• 5 random seeds are used.
• All test samples are evaluated for each seed.
• The average result is reported in the paper.
• *This setting follows SCCNet

We provide the weights of our model for direct use and reproduction.

• Download Model Weights
  https://github.com/XinnHe/HRL/releases/tag/weights

If you don’t feel like running HRL yourself, just leave me a message or drop me an email at [email protected]. I’ll be happy to share the HRL visualization results with you, based on your visualization style (blue, red, yellow, green, mask overlays, boundary highlighting, ......).

Let us engage in academic discussions on intelligent interpretation of remote sensing images.

@article{he2025hierarchical, title={Hierarchical Relation Learning for Few-shot Semantic Segmentation in Remote Sensing Images}, author={He, Xin and Liu, Yun and Zhou, Yong and Ding, Henghui and Zhao, Jiaqi and Liu, Bing and Jiang, Xudong}, journal={IEEE Transactions on Geoscience and Remote Sensing}, year={2025}, volume={63}, pages={4410615}, publisher={IEEE} }