Multi-LoRA Fine-Tuned Segment Anything Model for Extraction of Urban Man-Made Objects

Xiaoyan LU and Qihao WENG

[Paper]

Multi-LoRA Fine-Tuned SAM Framework

The training dataset

1. DeepGlobe Road Training Dataset : 4696 samples
2. SpaceNet Building AOI2 and AOI4 Dataset : 8429 samples

The validation set

1. DeepGlobe Road Test Dataset : 1530 samples
2. SpaceNet Building AOI3 and AOI5 Dataset : 1148 (Paris) and 1101 (Khartoum) samples
3. The WHU building (Christchurch) dataset: 2416 samples

2. The trained weights of SAM_Adapter, SAM_LoRA (r=96), and SAM_MLoRA (r=32,n=3) are released at Baidu Drive, Code: MODE

Road Extraction

SAM_Adapter

python train_sam_adapter.py --name='b_adapter_sam'

SAM_LoRA (r=96)

python train_sam_adapter.py --name='b_adapter_sam_lora96_96'

SAM_MLoRA (r=32,n=3)

python train_sam_adapter.py --name='b_adapter_sam_multi_lora'

Building Extraction

SAM_Adapter

python train_sam_adapter_build.py --name='b_adapter_sam_sp24'

SAM_LoRA (r=96)

python train_sam_adapter_build.py --name='b_adapter_sam_lora96_96_sp24'

SAM_MLoRA (r=32,n=3)

python train_sam_adapter_build.py --name='b_adapter_sam_multi_lora32_sp24'

Citation

If this code or dataset contributes to your research, please kindly consider citing our paper :)

@article{Lu2024MLoRA,
    title = {Multi-LoRA Fine-Tuned Segment Anything Model for Urban Man-Made Object Extraction},
    author = {Xiaoyan LU and Qihao Weng},
    journal = {IEEE Transactions on Geoscience and Remote Sensing},
    volume = {62},
    pages = {1-19},
    year = {2024},
    doi = {https://doi.org/10.1109/TGRS.2024.3435745}
}