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

Zeyu Ren1* Xiang Li2* Yiran Wang3* Zeyu Zhang2*† Hao Tang2‡
1 The University of Melbourne     2 Peking University     3 Australian Centre for Robotics
* Equal contribution.    † Project lead.    ‡ Corresponding author.

[Paper]      [Code]     [Dataset]     [Model]

TL;DR: StereoAdapter-2 achieves state-of-the-art underwater stereo depth estimation with a selective state space model.

Conceptual comparison explains how GRU’s complex gating contrasts with Selective SSM’s linear, input-dependent recurrence, inspiring ConvSS2D refinement.
Data synthesis pipeline. Semantic-aware style transfer and geometry-consistent novel view synthesis rendering pipeline for UW-StereoDepth-80K dataset.

Real World Results

Qualitative results of zero-shot underwater stereo depth estimation were obtained by deploying the model on a robotic platform.

Right Camera
Left Camera
Ground Truth
StereoAdapter-2

Simulation Results

Qualitative results of underwater stereo depth estimation on UW-StereoDepth-80K.

Right Camera
Left Camera
Ground Truth
StereoAdapter-2

Method

Detailed architecture of the StereoAdapter-2: Our model iteratively refines disparity by integrating a Mamba Adapter. The refinement step is powered by the ConvSS2D operator, which enables adaptive and long-range spatial information propagation through multi-directional selective scanning.

Benchmark Results

Qualitative results of zero-shot stereo depth estimation for different models on the SQUID dataset.



Qualitative results of zero-shot stereo depth estimation for different models on the TartanAir Ocean dataset

UW-StereoDepth-80K

Example from the high-quality UW-StereoDepth-80K dataset.

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