JunDong-dev/OptiSAR-Net-PlusPlus

OptiSAR-Net++ — Official Weights

Official trained weights (best.pt) of OptiSAR-Net++: A Large-Scale Benchmark and Transformer-Free Framework for Cross-Domain Remote Sensing Visual Grounding.

OptiSAR-Net++ is a transformer-free framework for Cross-Domain Remote Sensing Visual Grounding (CD-RSVG): a single unified model localizes targets described by free-form natural language in both optical and SAR remote sensing imagery, replacing a heavy Transformer decoder with contrastive region–text matching.

• 🤗 Dataset: JunDong-dev/OptSAR-RSVG
• 💻 Code: GitHub — JunDong-dev/OptiSAR-Net-PlusPlus

🏗️ Architecture

The checkpoint follows a YOLOE-style single-stage detector extended with:

Component	Location	Role
PLoRA-MoE	Backbone	Patch-level low-rank-adaptation Mixture-of-Experts for optical/SAR feature disentanglement
TGDG-SSA	Neck (×3)	Language-guided multi-scale fusion of visual features with text embeddings
OptiSARNetPlusPlusDetect	Head	Region–text contrastive matching head with region-aware auxiliary supervision
MobileCLIP2-B (frozen)	Text encoder	Encodes referring expressions

Key config of this checkpoint: nc: 16 classes, scale: m, reg_max: 16.

📦 Files

File	Description
best.pt	Trained weights (≈ 129 MB). The model architecture/modules are defined in the GitHub repository.

📈 Results — OptSAR-RSVG Test Split

Evaluated on the OptSAR-RSVG test split (4,434 images / 8,103 referring expressions). All values in %.

Domain	Samples	Pr@0.5	Pr@0.6	Pr@0.7	Pr@0.8	Pr@0.9	meanIoU	cumIoU
All	8,103	93.61	93.19	91.94	87.75	66.43	85.94	92.11
Optical	6,027	93.01	92.58	91.67	88.58	72.99	86.48	92.25
SAR	2,076	95.33	94.94	92.73	85.31	47.40	84.38	83.21

Benchmark comparisons against TransVG, LQVG, TACMT, CSDNet, Grounding DINO, GLIP, etc. are reported in the paper and the GitHub README.

🚀 Usage

best.pt contains custom modules (PLoRA_MoE, TGDG_SSA, OptiSARNetPlusPlusDetect), so it must be loaded with the project code rather than a stock Ultralytics install:

# 1. Clone the official repository (defines the custom modules & inference entry points)
git clone https://github.com/JunDong-dev/OptiSAR-Net-PlusPlus.git
cd OptiSAR-Net-PlusPlus
pip install -r requirements.txt

# 2. Download the checkpoint and the dataset
hf download JunDong-dev/OptiSAR-Net-PlusPlus best.pt --local-dir .
hf download JunDong-dev/OptSAR-RSVG --repo-type dataset --local-dir OptSAR-RSVG

# 3. Run evaluation / inference with the scripts provided in the repository README

Evaluation metrics follow the standard CD-RSVG protocol: Pr@{0.5–0.9}, meanIoU, cumIoU, with per-domain (optical / sar) reporting.

🎯 Intended Use

• Cross-domain (optical ↔ SAR) referring expression comprehension / visual grounding in remote sensing
• Research on multi-modal fusion, parameter-efficient MoE adaptation, and contrastive region–text matching

⚖️ License

Code and weights are released under AGPL-3.0. The companion dataset is subject to its source datasets' licenses (see the dataset card).

📚 Citation

@article{tang2026optisar,
  title={OptiSAR-Net++: A Large-Scale Benchmark and Transformer-Free Framework for Cross-Domain Remote Sensing Visual Grounding},
  author={Tang, Xiaoyu and Dong, Jun and Cheng, Jintao and Fan, Rui},
  journal={arXiv preprint arXiv:2603.24876},
  year={2026}
}