GeoBox-R1: Curriculum-Guided SFT and Geometric RL for Unified Box-Level Remote Sensing Visual Grounding

Chenxi Lan*, Yuchen Wu*, Minghang Zhou, Tianyu Li, Zhihao Qiu, Guoqing Wang^†

^*Equal contribution    ^†Corresponding author

Under review at AAAI 2027

Project page · Code

Overview

GeoBox-R1 is a 4B vision-language model for unified remote-sensing visual grounding. Given an aerial or satellite image and a referring expression, the same model can produce either a horizontal bounding box (HBB) or an oriented bounding box (OBB).

The model starts from Qwen3-VL-4B-Instruct and is trained in two stages:

1. Curriculum-guided SFT orders examples from HBB grounding to OBB grounding and then HBB-to-OBB chain-of-thought reasoning.
2. Geometric RL (GDPO) improves geometric precision with rotated-IoU and adaptive Wasserstein rewards, without a learned reward model.

Results

Macro averages are shown below. Full comparisons, per-dataset results, and the evaluation protocol are available on the project page.

Task	Evaluation sets	Acc@0.5	Acc@0.7	mIoU / mRIoU
HBB	7	58.78	42.22	50.39
OBB	3	47.32	27.55	39.85

Among the evaluated baselines, GeoBox-R1 achieves the best macro averages while using 4B parameters. GDPO is trained only on OBB samples, but it also improves HBB performance over the SFT stage.

Usage

Install a recent Transformers release together with PyTorch, Pillow, and Accelerate, then run:

from PIL import Image
from transformers import AutoModelForImageTextToText, AutoProcessor

model_id = "yuchenwu73/GeoBox-R1"

model = AutoModelForImageTextToText.from_pretrained(
    model_id,
    dtype="auto",
    device_map="auto",
)
processor = AutoProcessor.from_pretrained(model_id)

image = Image.open("scene.png").convert("RGB")
expression = "the brown SUV on the right"

prompt = f"""Locate the instance that matches the description: [{expression}]. Report oriented bbox coordinates in following JSON format:
```json
[
\t{{"oriented_bbox": [[x1, y1], [x2, y2], [x3, y3], [x4, y4]]}}
]
```"""

messages = [
    {
        "role": "user",
        "content": [
            {"type": "image"},
            {"type": "text", "text": prompt},
        ],
    }
]
text = processor.apply_chat_template(
    messages,
    tokenize=False,
    add_generation_prompt=True,
)
inputs = processor(text=[text], images=[image], return_tensors="pt").to(model.device)

generated = model.generate(**inputs, max_new_tokens=256)
generated = generated[:, inputs.input_ids.shape[1]:]
print(processor.batch_decode(generated, skip_special_tokens=True)[0])

For HBB grounding, replace the prompt with:

prompt = f"""Locate the instance that matches the description: [{expression}]. Report horizontal bbox coordinates in following JSON format:
```json
[
\t{{"horizontal_bbox": [x1, y1, x2, y2]}}
]
```"""

Coordinates are quantized to [0, 1000]. Multiply x coordinates by the image width divided by 1000, and y coordinates by the image height divided by 1000, to recover pixel coordinates.

The repository also provides evaluation scripts, an interactive demo, and the complete training pipeline: github.com/yuchenwu73/GeoBox-R1.

License

The model weights are released under the CC BY-NC 4.0 license. Users must also comply with the licenses and terms of the underlying Qwen3-VL model and any input datasets they use.

Citation

@misc{geoboxr1,
  title  = {GeoBox-R1: Curriculum-Guided SFT and Geometric RL for
            Unified Box-Level Remote Sensing Visual Grounding},
  author = {Lan, Chenxi and Wu, Yuchen and Zhou, Minghang and
            Li, Tianyu and Qiu, Zhihao and Wang, Guoqing},
  year   = {2026},
  url    = {https://yuchenwu73.github.io/geobox-r1/},
  note   = {Preprint}
}