RS-OVC: Open-Vocabulary Counting for Remote-Sensing Data

Pytorch Implementation for the paper RS-OVC: Open-Vocabulary Counting for Remote-Sensing Data - an open-vocabulary object-counting model for remote-sensing data, ICPR 2026.

Tamir Shor^1,2, George Leifman², Genady Beryozkin² ¹Technion – Israel Institute of Technology ²Google Research

Correspondence to: tamir.shor@campus.technion.ac.il

arXiv preprint: [pdf]

Overview

We present RS-OVC – the first open-vocabulary object counting framework for remote-sensing imagery. We show our model enables counting of novel object classes using textual and/or visual conditioning, without retraining.

Set-up

1. Clone repository

git clone https://github.com/tamirshor7/RS-OVC.git
cd RS-OVC

2. Create environment

conda create --name rsovc python=3.10
conda activate rsovc

Data

We curate a designated dataset from a set of common aerial imagery object counting and detection dataset, and adapt them for our task of novel-class counting.

Download Original Datasets

To replicate creation of our curated dataset used in the paper, download the following public dataset

NWPU-MOC https://github.com/lyongo/NWPU-MOC
FAIR-1M https://www.kaggle.com/code/ollypowell/fair1m-satellite-dataset-eda
DOTA https://captain-whu.github.io/DOTA/dataset.html
DIOR https://huggingface.co/datasets/torchgeo/dior
RSOC (Building class only) https://github.com/gaoguangshuai/Counting-from-Sky-A-Large-scale-Dataset-for-Remote-Sensing-Object-Counting-and-A-Benchmark-Method

Data Preprocessing

NWPU-MOC

python data/NWPU-MOC/coco_conversion.py --nwpu_root <NWPU_ROOT>/annotations/
python data/coco_to_odvg.py --data_dir data/NWPU-MOC/ --input_name nwpu_train_class_split.json

FAIR-1M

mkdir -p <FAIR1M_ROOT>/shared_images
cp <FAIR1M_ROOT>/Images/Train/* <FAIR1M_ROOT>/shared_images
cp <FAIR1M_ROOT>/Images/Val/* <FAIR1M_ROOT>/shared_images

python data/FAIR-1M/coco_conversion.py --fair1m_root <FAIR1M_ROOT>
python data/coco_to_odvg.py --data_dir data/FAIR-1M/ --input_name fair1m_train_class_split.json

DOTA

mkdir -p <DOTA_ROOT>/shared_images
cp <DOTA_ROOT>/train/images/images/* <DOTA_ROOT>/shared_images
cp <DOTA_ROOT>/val/images/images/* <DOTA_ROOT>/shared_images

python data/DOTA/coco_conversion.py --dota_root <DOTA_ROOT>
python data/coco_to_odvg.py --data_dir data/DOTA/ --input_name dota_train_class_split.json

DIOR

mkdir -p <DIOR_ROOT>/shared_images
cp <DIOR_ROOT>/JPEGImages-trainval/* <DIOR_ROOT>/shared_images
cp <DIOR_ROOT>/JPEGImages-test/* <DIOR_ROOT>/shared_images

python data/DIOR/coco_conversion.py --dior_root <DIOR_ROOT>
python data/coco_to_odvg.py --data_dir data/DIOR/ --input_name dior_train_class_split.json

RSOC (Building)

python data/RSOC-Building/coco_conversion.py --rsocb_root <RSOCB_ROOT>
python data/coco_to_odvg.py --data_dir data/RSOC-Building/ --input_name rsocb_train_class_split.json

Pretrained Weights (CountGD Initialization)

mkdir checkpoints

Download:

https://drive.google.com/file/d/1RbRcNLsOfeEbx6u39pBehqsgQiexHHrI

Place at:

checkpoints/checkpoint_fsc147_best.pth

Configure Dataset Paths

Update dataset paths inside:

config/datasets_*.json

Example (Linux):

# NWPU-MOC
sed -i -E 's|("root":\s*")data/NWPU-MOC|<NWPU_ROOT>|g' config/*.json

# FAIR-1M
sed -i -E 's|("root":\s*")data/FAIR-1M|<FAIR1M_ROOT>|g' config/*.json

# DIOR
sed -i -E 's|("root":\s*")data/DIOR|<DIOR_ROOT>|g' config/*.json

# DOTA
sed -i -E 's|("root":\s*")data/DOTA|<DOTA_ROOT>|g' config/*.json

# RSOC
sed -i -E 's|("root":\s*")\.\./data/RSOC_building|<RSOCB_ROOT>|g' config/*.json

Training

RS-OVC (main model)

torchrun --standalone --nnodes=1 --nproc-per-node=4 \
main.py \
--config_file config/cfg_fndd.py \
--datasets config/datasets_shared_fndd.json \
--mode fused

If you encounter NCCL issues:

NCCL_P2P_DISABLE=1 NCCL_IB_DISABLE=1 torchrun ...

Single-GPU training also works with minimal performance degradation.

Baselines

Mode	Description
`fused`	RS-OVC (ours)
`rs_only`	RS encoder only
`rsft`	RS finetuning
`countgd`	Original CountGD

to train any of the baselines from the paper, set the mode argument to either of ‘rs_only’, ‘rsft’,’countgd’ (naming matches naming conventions from the paper). To train in RSFT mode pass –config_file config/cfg_fndd_rsft.py (additionaly to –mode rsft). To train in RS-Only mode pass –config_file config/cfg_fndd_rs_only.py (additionaly to –mode rs_only).

Examples:

# RS-FT
python main.py --config_file config/cfg_fndd_rsft.py --mode rsft ...

# RS-Only
python main.py --config_file config/cfg_fndd_rs_only.py --mode rs_only ...

Evaluation

python main.py \
--config_file config/cfg_<dataset>_<mode>.py \
--datasets config/datasets_<dataset>.json \
--mode fused \
--eval \
--resume <CHECKPOINT_PATH>

Example (NWPU-MOC)

python main.py \
--config_file config/cfg_nwpu_fused.py \
--datasets config/datasets_nwpu.json \
--mode fused \
--eval \
--resume <CHECKPOINT_PATH>

Checkpoints

RS-OVC trained checkpoints:

wget -O checkpoint.pth https://huggingface.co/tamirshor/RSOVC/resolve/main/checkpoint.pth

Acknowledgments

This work builds upon CountGD. Our method extends its architecture and training framework, and parts of this codebase reuse and adapt components from the original implementation. We thank the authors for publicly-releasing their code and enabling our work to build-upon theirs.

Citation

@article{shor2026rs,
  title={RS-OVC: Open-Vocabulary Counting for Remote-Sensing Data},
  author={Shor, Tamir and Leifman, George and Beryozkin, Genady},
  journal={arXiv preprint arXiv:2604.08704},
  year={2026}
}