StereoAdapter-2: Globally Structure-Consistent Underwater Stereo Depth Estimation

This is the official repository for the paper:

StereoAdapter-2: Globally Structure-Consistent Underwater Stereo Depth Estimation

Zeyu Ren^1*, Xiang Li^2*, Yiran Wang^3*, Zeyu Zhang^2*†, and Hao Tang^2‡

¹The University of Melbourne, ²Peking University, ³Australian Centre for Robotics

^*Equal contribution. ^†Project lead. ^‡Corresponding author: bjdxtanghao@gmail.com.

Paper | Website | Dataset | Model

Introduction

Conceptual comparison explains how GRU's complex gating contrasts with Selective SSM's linear, input-dependent recurrence, inspiring ConvSS2D refinement.

✏️ Citation

If you find our code or paper helpful, please consider starring ⭐ us and citing:

@misc{ren2026stereoadapter2globallystructureconsistentunderwater,
  title={StereoAdapter-2: Globally Structure-Consistent Underwater Stereo Depth Estimation},
  author={Zeyu Ren and Xiang Li and Yiran Wang and Zeyu Zhang and Hao Tang},
  year={2026},
  eprint={2602.16915},
  archivePrefix={arXiv},
  primaryClass={cs.CV},
  url={https://arxiv.org/abs/2602.16915},
}

🔧 Quick Start

1. Install & Requirements

conda env create -f environment.yaml
conda activate stereoadapter2

2. Compile CUDA Kernels

Follow the installation instructions from Mamba and VMamba to compile the required CUDA kernels.

3. Download Pretrained Depth Anything 3 Model

Download the pretrained weights from Depth Anything 3 or directly from HuggingFace:

mkdir -p depthanything/checkpoints
cd depthanything/checkpoints
hf download depth-anything/DA3-BASE --local-dir DA3-BASE
cd ../..

4. Train

bash train_stereo_da3_codyra_encoder_mono_ss2d_decoder_without_inp_list.sh

😘 Acknowledgement

• RAFT-Stereo
• Depth Anything 3
• VMamba
• StereoAdapter

📜 License

This work is licensed under CC BY-NC-SA 4.0.