SurroundNEXO: Ego-Centric Metric Bridging for Spatially Consistent Geometry in Autonomous Driving

   Shuai Yuan,^1,2,*   Runxi Tang,¹   Yuzhou Ji,¹   Fudong Ge,^1,2   Hanshi Wang,^1,2  
Yifei Wang,²   Xianming Zeng,²   Jianyun Xu,²   Xinglinag Liu,²   Yanfeng Wang,¹   Zhipeng Zhang^{1 ✉}  

¹School of Artificial Intelligence, Shanghai Jiao Tong University  
²Hello Inc.

^✉ Corresponding Author

SurroundNEXO is an ego-centric metric depth framework tailored for low-overlap, surround-view autonomous driving scenes. It bridges weakly overlapping cameras through (i) ego-ray positional encoding for a shared geometric reference, (ii) sparse metric anchoring for absolute scale propagation, and (iii) a progressive geometry transformer for stable view-local, cross-view, and global interaction — all within a unified network.

📰 News

[2026-06] SurroundNEXO paper is released on arXiv!

[2026-06] SurroundNEXO inference code and model are released.

📖 Overview

$\spadesuit$ We propose SurroundNEXO, a novel feed-forward framework specifically designed to achieve metric-scale accurate and spatially consistent cross-view depth perception in low-overlap autonomous driving scenes.

$\spadesuit$ We introduce ego-ray positional encoding (ERPE) to provide a shared spatial parameterization for view consistency, sparse metric anchoring (SMA) to establish stable metric anchors for absolute depth accuracy, and progressive geometry transformer (PGT) to facilitate highly efficient, coarse-to-fine feature interaction.

$\spadesuit$ Extensive experiments demonstrate that our method achieves state-of-the-art performance across standard benchmarks, significantly improving single-frame accuracy and cross-view consistency.

🌍 Installation

1. Clone SurroundNEXO

git clone https://github.com/AutoLab-SAI-SJTU/SurroundNEXO.git
cd SurroundNEXO

2. Create conda environment

conda create -n surroundnexo python=3.10
conda activate surroundnexo

3. Install requirements

pip install -r requirement.txt

🤗 Checkpoints

We provide the pretrained SurroundNEXO weights through Hugging Face. The inference code downloads model.safetensors from AutoLab-SJTU/SurroundNEXO with huggingface_hub.

pip install -U huggingface_hub
cd ./SurroundNEXO
mkdir -p ckpt

hf download AutoLab-SJTU/SurroundNEXO model.safetensors \
    --local-dir ./ckpt

▶️ Run Inference

# Run the provided NuScenes example
python inference.py \
    --checkpoint ./ckpt/model.safetensors \
    --input_path examples/nuscenes-002 \
    --output_dir ./output_surroundnexo 

📋 Checklist

• [ √ ] Release the pre-trained checkpoints for SurroundNEXO.
• [ √ ] Release the inference code.
• [ ] Release the evaluation code.
• [ ] Release the training and data-processing code.

🙏 Acknowledgement

We would like to acknowledge the following open-source projects that served as a foundation for our implementation:

Depth-Anything-3   DVGT   MoGe-2

Many thanks to these authors!

📜 Citation

If you incorporate our work into your research, please cite:

@article{yuan2026surroundnexo,
        title   = {SurroundNEXO: Ego-Centric Metric Bridging for Spatially Consistent Geometry in Autonomous Driving},
        author  = {Yuan, Shuai and Tang, Runxi and Ji, Yuzhou and Ge, Fudong and Wang, Hanshi and Wang, Yifei and Zeng, Xianming and Xu, Jianyun and Liu, Xingliang and Wang, Yanfeng and Zhang, Zhipeng},
        journal = {arXiv preprint arXiv:2606.16960},
        year    = {2026}
}