Urban Segmentation is a streamlined, high-performance framework designed for semantic segmentation of remote sensing imagery.
Leveraging the power of UNICOM—a vision foundation model pre-trained on 400 million images—this project demonstrates extreme data efficiency. We achieve SOTA-level segmentation results using only 4 labeled satellite images for training. This repository serves as both a robust baseline for research and an accessible entry point for practitioners.
Urban Segmentation 是一个专为遥感图像语义分割设计的高效框架。
本项目利用了在 4亿 海量数据上预训练的视觉基础模型 UNICOM，展示了极致的数据样本效率。我们仅需 4张 标注的卫星图像进行微调，即可获得极佳的分割效果。这不仅为科研提供了一个强有力的 Baseline，也为初学者提供了一个极简的实战范例。
- Foundation Model Power: Built upon UNICOM ViT backbones, inheriting robust feature representations.
- Extreme Few-Shot: Achieve high mIoU with minimal annotated data (4 images).
- Plug-and-Play: Minimalist code structure (~200 lines for training) without complex dependencies.
- Scalable: Supports both single-GPU rapid prototyping and multi-GPU distributed training.
| Dynamic Prediction | Generalization Test |
|---|---|
 |
 |
 |
 |
git clone https://github.com/anxiangsir/urban_seg.git
cd urban_seg
pip install -r requirements.txtDownload the dataset (CCF Satellite Imagery) from Baidu Cloud (Code: 3ih2).
Structure your directory as follows:
dataset/
├── origin/ # 5 annotated source images
├── test/ # Unlabeled test images
└── train/ # Generated via preprocessing
├── images/
└── labels/
Run the preprocessing script to generate random crops:
python preprocess.pyDownload the pre-trained UNICOM weights from the Official Release:
FP16-ViT-B-32.ptFP16-ViT-B-16.pt(Recommended)FP16-ViT-L-14.pt
Option A: Rapid Prototyping (Single GPU)
# Minimal implementation (~200 lines)
python train_one_gpu.pyOption B: High-Performance Training (Multi-GPU DDP)
# Distributed Data Parallel
torchrun --nproc_per_node 8 train_multi_gpus.pyIf you find this project or the UNICOM model useful for your research, please consider citing:
@inproceedings{an2023unicom,
title={Unicom: Universal and Compact Representation Learning for Image Retrieval},
author={An, Xiang and Deng, Jiankang and Yang, Kaicheng and Li, Jiawei and Feng, Ziyong and Guo, Jia and Yang, Jing and Liu, Tongliang},
booktitle={ICLR},
year={2023}
}
We welcome all contributions! Please feel free to open an issue or submit a pull request.