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

Zeyu Ren; Xiang Li; Yiran Wang; Zeyu Zhang; Hao Tang

cs.CVFeb 18, 2026

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

Authors: Zeyu Ren, Xiang Li, Yiran Wang, Zeyu Zhang, Hao Tang

Organizations: The University of Melbourne · Peking University · Australian Centre for Robotics

Abstract

Stereo depth estimation is fundamental to underwater robotic perception, yet suffers from severe domain shifts caused by wavelength-dependent light attenuation, scattering, and refraction. Recent approaches leverage monocular foundation models with GRU-based iterative refinement for underwater adaptation; however, the sequential gating and local convolutional kernels in GRUs necessitate multiple iterations for long-range disparity propagation, limiting performance in large-disparity and textureless underwater regions. In this paper, we propose StereoAdapter-2, which replaces the conventional ConvGRU updater with a novel ConvSS2D operator based on selective state space models. The proposed operator employs a four-directional scanning strategy that naturally aligns with epipolar geometry while capturing vertical structural consistency, enabling efficient long-range spatial propagation within a single update step at linear computational complexity. Furthermore, we construct UW-StereoDepth-80K, a large-scale synthetic underwater stereo dataset featuring diverse baselines, attenuation coefficients, and scattering parameters through a two-stage generative pipeline combining semantic-aware style transfer and geometry-consistent novel view synthesis. Combined with dynamic LoRA adaptation inherited from StereoAdapter, our framework achieves state-of-the-art zero-shot performance on underwater benchmarks with 17% improvement on TartanAir-UW and 7.2% improvment on SQUID, with real-world validation on the BlueROV2 platform demonstrates the robustness of our approach. Code: https://github.com/AIGeeksGroup/StereoAdapter-2. Website: https://aigeeksgroup.github.io/StereoAdapter-2.

Figures & tables

Fig. 1 : Conceptual comparison. The Gated Recurrent Unit (GRU) relies on multiple non-linear gates and candidate states

\tilde{h}_{t}

to update the hidden state

{h}_{t}

. Its complex gating mechanism introduces non-linear recursion that is difficult to analyze for long sequences. The Selective SSM streamlines this into a linear recurrence. By dynamically generating parameters from the input

x_{t}

, the Selective SSM maintains ”input-dependent selectivity” to adaptively modulate information flow. We leveraged the characteristics of selective SSM to design ConvSS2D, enabling the adaptation iterative process.

Fig. 2 : 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.

Fig. 3 : Data synthesis pipeline. Semantic-aware style transfer and geometry-consistent novel view synthesis rendering pipeline for UW-StereoDepth-80K dataset.

Fig. 4 : Qualitative results of zero-shot stereo depth estimation

Method	Training Set	Rel $\downarrow$	SqRel $\downarrow$	RMSE $\downarrow$	Log RMSE $\downarrow$	A1 $\uparrow$	A2 $\uparrow$	A3 $\uparrow$
LEAStereo [ 11 ]	Scene Flow	0.1099	1.3898	4.5610	0.2063	0.8929	0.9512	0.9761
PSMNet [ 8 ]	Scene Flow	0.0884	0.8699	3.9721	0.1804	0.9122	0.9627	0.9804
AANet [ 68 ]	Scene Flow	0.6096	8.3687	13.0542	0.9903	0.2598	0.3451	0.3888
GwcNet [ 21 ]	Scene Flow	0.1013	1.2965	4.1829	0.1855	0.9085	0.9612	0.9801
ACVNet [ 65 ]	Scene Flow	0.0970	1.1335	3.9985	0.1803	0.9063	0.9612	0.9813
RAFT-Stereo [ 37 ]	Scene Flow	0.0814	0.7342	4.0423	0.1703	0.9030	0.9612	0.9832

TABLE I : Quantitative comparison of zero-shot stereo depth estimation on the TartanAir underwater subset. All methods are evaluated under the same protocol using standard depth metrics.

Fig. 5 : Qualitative results of zero-shot underwater stereo depth estimation were obtained by deploying the model on a robotic platform.

Method	Training Set	Rel $\downarrow$	SqRel $\downarrow$	RMSE $\downarrow$	Log RMSE $\downarrow$	A1 $\uparrow$	A2 $\uparrow$	A3 $\uparrow$
LEAStereo [ 11 ]	Scene Flow	0.5574	3.9434	5.4659	0.4335	0.6512	0.8042	0.8869
PSMNet [ 8 ]	Scene Flow	0.5182	7.1404	4.9186	0.5902	0.7139	0.7999	0.8311
AANet [ 68 ]	Scene Flow	7.4801	314.1577	34.7612	1.8994	0.0602	0.1087	0.1570
GwcNet [ 21 ]	Scene Flow	0.2294	1.2275	3.0003	0.3799	0.7423	0.8517	0.9005
ACVNet [ 65 ]	Scene Flow	1.6030	65.6518	10.3828	0.7293	0.7019	0.7925	0.8321
RAFT-Stereo [ 37 ]	Scene Flow	0.0831	0.6946	1.9625	0.1441	0.9235	0.9634	0.9835

TABLE II : Zero-shot evaluation on SQUID dataset. 5 Datasets ∗ refers to Scene Flow [ 44 ] , Sintel [ 7 ] , ETH3D [ 49 ] , InStereo2K [ 3 ] , and CREStereo [ 30 ] .

Method	REL $\downarrow$	SqRel $\downarrow$	RMSE $\downarrow$	Log RMSE $\downarrow$	A1 $\uparrow$
Stereo Anywhere [ 4 ]	0.1218	1.0623	2.4682	0.1673	0.8541
FoundationStereo [ 61 ]	0.1304	0.6187	2.0893	0.1635	0.8812
StereoAdapter [ 62 ]	0.1163	0.6794	1.9285	0.1556	0.8694
StereoAdapter-2 (Ours)	0.1023	0.5843	1.7164	0.1354	0.9256

TABLE III : Real-world evaluation on BlueROV2.

Fig. 6 : Hardware platform for real world experiments.

DA3 Encoder	Mono Disp. Init.	Context Encoder	Update Module	REL $\downarrow$	RMSE $\downarrow$
		✓	ConvGRU	0.0516	2.82
✓		✓	ConvGRU	0.0482	2.64
✓		✓	ConvSS2D	0.0449	2.46
✓			ConvSS2D	0.0463	2.54
✓	✓		ConvSS2D	0.0440	2.40

TABLE IV : Model ablation of StereoAdapter-2, evaluating the effects of different design components, including the Depth Anything 3 encoder, monocular disparity initialization, context encoder, and update module.

Batch Size	Learning Rate	Train Iters	REL $\downarrow$	RMSE $\downarrow$
4	$1\times 10^{-4}$	16	0.0461	2.53
4	$2\times 10^{-4}$	16	0.0489	2.68
8	$1\times 10^{-4}$	16	0.0453	2.47
8	$2\times 10^{-4}$	16	0.0476	2.59
8	$1\times 10^{-4}$	22	0.0440	2.40

TABLE V : Ablation on training hyperparameters.

d_state	ssm_ratio	Params	FLOPs	TP.	REL $\downarrow$	RMSE $\downarrow$
		(M)	(G)	(img/s)
1	1.0	20.40	843.33	5.26	0.0445	2.42
4	1.0	20.42	845.81	5.22	0.0440	2.40
16	1.0	20.47	855.72	4.71	0.0438	2.38
16	1.5	20.60	886.17	4.62	0.4430	2.43
16	2.0	20.73	916.62	4.26	0.4551	2.45

TABLE VI : Ablation study on ConvSS2D SSM hyperparameters under FP32 precision, analyzing the effects of the state dimension

d_{\text{state}}

and SSM ratio on model accuracy and efficiency.

Scanning Pattern	Params	FLOPs	TP.	REL $\downarrow$	RMSE $\downarrow$
	(M)	(G)	(img/s)
Unidi-Scan	20.47	855.72	4.87	0.0459	2.46
Bidi-Scan	20.47	855.72	4.86	0.0453	2.42
Cross-Scan	20.47	855.72	4.74	0.0440	2.40

TABLE VII : Ablation on SS2D scanning patterns.

Method	Params (M)	On-board (ms)
FoundationStereo [ 61 ]	375	1933
Stereo Anywhere [ 4 ]	347	1524
MGStereo [ 5 ]	347	1631
StereoAdapter [ 62 ]	202	1285
StereoAdapter-2 (Ours)	103	1102

TABLE VIII : Average per-frame inference latency (ms) on Jetson Orin NX @

640{\times}360

, batch size=1.

Appendix figures & tables3 assets

Supplementary material from the paper’s appendix.

Appendix

Fig. 10 : Visualization of UW-StereoDepth-80K

Depth Estimation

Depth estimation infers scene distance from images and sensors, enabling 3D reconstruction, navigation, and spatial understanding.

See all 198 papers

AquaStereo: Enabling Underwater Stereo Matching via Depth-Conditioned Diffusion and Geometry Self-Distillation
Jul 5, 2026Qizhe Wei, Yingping Liang, Shaodi You +1Underwater Imaging Stereo Matching
Beijing Institute of Technology, Beijing, China · University of Amsterdam, Amsterdam, The Netherlands
SurfSLAM: Sim-to-Real Underwater Stereo Reconstruction For Real-Time SLAM
Jan 15, 2026Onur Bagoren, Seth Isaacson, Sacchin Sundar +6Simultaneous Localization and Mapping Metric Depth Estimation
Department of Robotics, University of Michigan, Ann Arbor, MI 48109, USA
LinStereo: Linear-Complexity Global Attention for Multi-Scale Iterative Stereo Matching
Jun 24, 2026Yiran Wang, Oliver Turner, Viorela IlaMulti-Scale Feature Fusion Stereo Matching
Australian Centre for Robotics (ACFR), School of Aerospace, Mechanical and Mechatronic Engineering (AMME), Faculty of Engineering, The University of Sydney, Sydney, NSW, Australia

Jul 5, 2026J/K move · Enter open · S save

AquaStereo: Enabling Underwater Stereo Matching via Depth-Conditioned Diffusion and Geometry Self-Distillation

Qizhe Wei, Yingping Liang, Shaodi You, Ying Fu

Beijing Institute of Technology, Beijing, China · University of Amsterdam, Amsterdam, The Netherlands

Underwater Imaging Stereo Matching Stereo Depth Estimation

Open paper AskPDF arXiv

Learning-based stereo matching models struggle in underwater environments due to scarce in-domain data and the difficulty of extracting discriminative correspondences from degraded imagery. In this work, we present

\textbf{AquaStereo}

, a perception-enhanced framework with a data simulation pipeline and a self-distillation strategy that jointly address data scarcity and feature degradation in underwater stereo matching. First, a depth-conditioned diffusion pipeline renders underwater stereo pairs while preserving binocular geometry, with a lightweight left-right consistency module ensuring geometric alignment. Training on this synthetic corpus effectively narrows the terrestrial-underwater gap and improves zero-shot robustness. Second, a frozen binocular teacher trained on clean terrestrial pairs guides a student exposed to rendered underwater pairs with perturbations. A stage-weighted sequence loss is performed to align the student's disparities with the teacher's geometry, while a clean-branch supervision with shared pseudo targets prevents scale drift. To further enhance feature stability under turbidity and low texture, we introduce learnable perception frames, a perception-enhanced feature formulation that constructs robust matching descriptors by fusing temporal cues from two auxiliary views encoded by a video backbone with semantic features extracted by a strong image encoder. Extensive experiments demonstrate that

\textbf{AquaStereo}

substantially improves robustness and zero-shot generalization in challenging underwater scenarios. The code is available at https://github.com/qz-wei/AquaStereo.

Abstract

Depth Estimation

Explore similar work

AquaStereo: Enabling Underwater Stereo Matching via Depth-Conditioned Diffusion and Geometry Self-Distillation

SurfSLAM: Sim-to-Real Underwater Stereo Reconstruction For Real-Time SLAM

LinStereo: Linear-Complexity Global Attention for Multi-Scale Iterative Stereo Matching