HSLabeling: Towards Efficient Labeling for Large-scale Remote Sensing Image Segmentation with Hybrid Sparse Labeling (HSLabeling)

Code for TIP 2025 paper, ["HSLabeling: Towards Efficient Labeling for Large-scale Remote Sensing Image Segmentation with Hybrid Sparse Labeling"] Authors: Jiaxing Lin, Zhen Yang, Qiang Liu, Yinglong Yan, Pedram Ghamisi, Weiying Xie, and Leyuan Fang

Getting Started

Prepare Dataset

Download the Potsdam and Vaihingen datasets after processing.

you can download the datasets from the official website and link. Then, crop the original images and create labels following our code in Dataprocess.

If your want to run our code on your own datasets, the pre-process code is also available in Dataprocess.

Evaluate

1. Download the Potsdam and Vaihingen datasets and LoveDA datasets

2. Download our weight

3. Run our code

python predict.py

Train

1. Generate SAM labels

python segment-anything/notebooks/automatic_mask_generator.py

2. Generate prospective scribble and block labels base on SAM labels

python DataProcess/sparse_label_generator.py

3. Generate hybrid sparse labels and Train segmentation model

python run/point/generate_train.py.py

Acknowledgement

We thank DBFNet and Segment Anything for part of their codes, processed datasets, data partitions, and pretrained models.