ZoRI: Towards Discriminative Zero-Shot Remote Sensing Instance Segmentation

This repository contains code for our AAAI2025 paper: ZoRI: Towards Discriminative Zero-Shot Remote Sensing Instance Segmentation.

TODO List ✅

• Modules implementation
• Dataset release
• Main code release

Installation

Please see Installation Instructions.

Datasets

Download datasets for zero-shot remote sensing instance segmentation from One Drive ☁️.

iSAID, NWPU, SIOR.

Download links will be updated soon...

Getting Started

Please see Getting Started with ZoRI.

Training

python train_net.py --config-file configs/zori_isaid_11_4.yaml

Inference

For GZSRI setting, run

python train_net.py  --config-file configs/zori_isaid_11_4.yaml --eval-only MODEL.WEIGHTS [path_to_weights]

For ZSRI setting, run

python train_net.py  --config-file configs/zori_isaid_11_4.yaml --eval-only MODEL.WEIGHTS [path_to_weights] DATASETS.TEST '("isaid_zsi_11_4_val_unseen",)' MODEL.GENERALIZED False MODEL.CACHE_BANK.ALPHA 0.6

Finally, inference again with pseudo unseen visual prototypes to get final predictions.

Acknowledgement

This project is based on FC-CLIP. Many thanks to the authors for their great work!

BibTeX

Please consider to cite ZoRI if it helps your research.

@inproceedings{ZoRI,
title={ZoRI: Towards Discriminative Zero-Shot Remote Sensing Instance Segmentation},
author={Huang, Shiqi and He, Shuting and Wen, Bihan},
booktitle={Proceedings of the AAAI Conference on Artificial Intelligence},
year={2025}
}