PaperlayerView exact revision on arXiv

Source-linked AI summary

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

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

arXiv:2602.16915v1cs.CV

TL;DR

Underwater stereo depth estimation is limited by optical domain shifts, difficult long-range disparity refinement, and scarce diverse ground-truth data. StereoAdapter-2 combines ConvSS2D selective state space updates, four-directional scanning, UW-StereoDepth-80K, and dynamic LoRA adaptation. It reports state-of-the-art zero-shot performance, including 17% improvement on TartanAir-UW and 7.2% on SQUID, with BlueROV2 validation.

  • Problem

    Underwater stereo estimation lacks robust long-range refinement and diverse accurately annotated data under attenuation, scattering, and refraction.

  • Method

    StereoAdapter-2 replaces ConvGRU with four-directional ConvSS2D selective state space updates and constructs UW-StereoDepth-80K for dynamic LoRA adaptation.

  • Results

    17% improvement on TartanAir-UW and 7.2% on SQUID are reported alongside state-of-the-art zero-shot benchmark performance and BlueROV2 validation.

  • Takeaways & Limitations

    The framework supports robust generalization from synthetic training to real underwater scenes and efficient long-range spatial propagation.

  • Takeaways & Limitations

    The synthetic-to-real gap persists under severe turbidity, strong backscatter, or rapidly varying illumination, and temporal consistency remains challenging during continuous deployment.

Abstract

from arXiv · show

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.

I. INTRODUCTION

StereoAdapter-2 addresses underwater stereo depth estimation through a selective-state-space update operator and a large synthetic dataset, targeting long-range disparity refinement and synthetic-to-real diversity. Combined with dynamic LoRA adaptation, it reports state-of-the-art zero-shot benchmark performance and BlueROV2 validation.

  • Motivation: Underwater stereo depth estimation faces domain shifts from wavelength-dependent attenuation, scattering, and refraction, while accurate depth remains important for robotic perception.These conditions complicate reliable stereo matching and affect underwater autonomous operations.
  • Motivation: Two practical challenges are improving iterative disparity refinement in large-disparity and textureless regions and narrowing the synthetic-to-real gap caused by scarce annotated underwater stereo data.The paper frames these as complementary architectural and data problems.
  • Results: 17% improvement on TartanAir-UW and 7.2% on SQUID accompany state-of-the-art zero-shot underwater benchmark performance and BlueROV2 real-world validation.The framework combines dynamic LoRA adaptation with ConvSS2D and UW-StereoDepth-80K.
  • Method: ConvSS2D replaces ConvGRU with selective state space models and four-directional scanning that captures horizontal epipolar constraints and vertical structural consistency.The design enables long-range spatial propagation within a single refinement step at linear computational complexity.
  • Data: UW-StereoDepth-80K is a large-scale synthetic underwater stereo dataset with diverse baselines and optical parameters generated through a two-stage pipeline.The pipeline combines semantic-aware style transfer with geometry-consistent novel view synthesis.

II. RELATED WORK

Underwater stereo matching is challenged by unreliable supervision, distorted cross-view appearance, and limited scene diversity. State space models and multi-directional scanning motivate efficient long-range refinement suited to stereo geometry.

  • Underwater depth estimation and datasets: Accurate dense underwater disparity labels are difficult because LiDAR is unreliable underwater and large-scale data collection is costly.
  • Underwater depth estimation and datasets: FLSea-Stereo lacks accurate stereo-disparity annotations, while UWStereo remains less complex than real-world underwater scenarios.
  • Underwater depth estimation and datasets: Light scattering, absorption, and refraction reduce photometric consistency between underwater views, making reliable matching difficult.
  • State Space Model: State-space models efficiently capture long-range dependencies with linear or near-linear sequence complexity, while selective scanning improves long-sequence modeling.
  • State Space Model: SS2D scans images along multiple spatial directions to capture spatial dependencies, motivating its use for stereo matching alongside LoRA-based domain adaptation.
  • State Space Model: After discretization, SSMs support linear recurrence and global convolution, with the structured convolutional kernel later reinterpreted for spatial disparity refinement.

A. Overview

StereoAdapter-2 combines a monocular depth foundation model with parameter-efficient adaptation and stereo-specific correlation features. A multi-scale correlation pyramid supplies matching evidence across fine and large displacements.

  • Overview: StereoAdapter-2 uses Depth Anything 3 as both feature encoder and monocular depth estimator, with LoRA enabling efficient underwater adaptation.
  • Overview: Monocular depth estimation initializes disparity to accelerate convergence, while encoder fine-tuning follows the StereoAdapter adaptation approach.
  • Overview: Features are extracted from four intermediate Transformer layers to capture multi-scale details and semantic information.
  • Overview: The correlation volume computes inner products between left and right features sharing the same y coordinate, reflecting calibrated horizontal stereo correspondence.
  • Overview: A four-layer correlation pyramid uses average pooling along the right-image dimension to provide progressively larger receptive fields at fixed spatial resolution.
  • Overview: Each refinement iteration linearly interpolates correlation values around the current disparity across pyramid layers and concatenates them as update inputs.

D. Iterative Disparity Estimation

StereoAdapter-2 replaces ConvGRU refinement with ConvSS2D, using sequential state recursion and four-directional scans for long-range spatial propagation. Input-dependent parameters further adapt state updates to image content.

  • D. Iterative Disparity Estimation: StereoAdapter-2 iteratively updates disparity from an initial estimate, replacing ConvGRU with ConvSS2D as the core refinement operator.
  • D. Iterative Disparity Estimation: ConvSS2D models long-range dependencies through sequential state recursion without requiring multiple convolutional layers to expand the receptive field.
  • D. Iterative Disparity Estimation: The hidden state propagates information from previous spatial positions, allowing distant locations to influence one another within a single refinement step.
  • D. Iterative Disparity Estimation: ConvSS2D discards the traditional context encoder and directly projects decoder features to initialize the hidden state.
  • D. Iterative Disparity Estimation: Unlike ConvGRU’s predominantly local aggregation, ConvSS2D computes input-dependent selectivity through dynamically generated parameters ∆, B, and C.
  • D. Iterative Disparity Estimation: The parameters ∆, B, and C respectively modulate state evolution, feature incorporation, and output projection based on local image characteristics.
  • D. Iterative Disparity Estimation: Four-directional scanning extends selective scanning across horizontal and vertical directions, aligning horizontal propagation with epipolar constraints while aggregating two-dimensional context.

E. Data Synthesis: UW-StereoDepth-80K

UW-StereoDepth-80K is built through a two-stage pipeline that transforms terrestrial RGB-D data into underwater stereo pairs with varied appearance and camera geometry. The resulting dataset provides synthetic underwater imagery, consistent stereo geometry, and dense disparity supervision.

  • Data Synthesis: The pipeline sequentially applies semantic-aware style transfer and geometry-consistent novel view synthesis to terrestrial RGB-D data.Atlantis provides underwater appearance transfer, while NVS-Solver generates the stereo view using camera geometry.
  • Data Synthesis: Semantic-aware style transfer hallucinates attenuation, scattering, and turbidity while preserving scene semantics and geometric structure.Depth-conditioned diffusion transforms terrestrial images into underwater imagery without losing geometric labels.
  • Data Synthesis: NVS-Solver synthesizes the target stereo view by conditioning novel-view generation on explicit baseline displacements.The method treats stereo generation as a view-synthesis problem to maintain multi-view geometric consistency.
  • Dataset Construction: The generated 640 × 480 stereo pairs combine physically plausible underwater appearance, consistent stereo geometry, and dense ground-truth disparity.The dataset merges diffusion-based samples with the existing UW-StereoDepth-40K dataset.
  • Dataset Construction: UW-StereoDepth-80K contains about 80K samples synthesized from virtual underwater data and evaluated against TartanAir-UW and SQUID.TartanAir-UW contains 13,583 underwater stereo pairs, while SQUID contains four distinct scenes.

B. Implementation Details

StereoAdapter-2 is evaluated with a fixed training and inference setup and achieves strong zero-shot results on underwater benchmarks. Its performance is reported against standard baselines and prior StereoAdapter variants.

  • Implementation Details: StereoAdapter-2 uses 22 training iterations and 32 inference iterations with Depth Anything 3 ViT-B initialization and inherited LoRA settings.The LoRA configuration uses rank r = 16, sparsity threshold κmax = 0.005, and regularization weight λ = 1 × 10^-4.
  • Benchmark Results: StereoAdapter-2 achieves state-of-the-art zero-shot performance across TartanAir Underwater and SQUID without target-domain fine-tuning.The approach consistently outperforms existing stereo matching methods under the reported evaluation protocol.
  • Benchmark Results: 16.5% lower REL and 17.0% lower RMSE are reported on TartanAir Underwater versus StereoAdapter trained on UW-StereoDepth-40K.StereoAdapter-2 reports REL 0.0440, RMSE 2.4038, and A1 accuracy 96.76%.
  • Benchmark Results: 7.2% lower RMSE is reported on SQUID versus the previous StereoAdapter, alongside REL 0.0705 and RMSE 1.7481.The reported threshold accuracies are A1 94.25%, A2 97.65%, and A3 98.62%.
  • Qualitative Results: Qualitative zero-shot results show more accurate and visually coherent depth maps, including better scale estimation for far-range details.The comparison is presented against baseline methods in Figure 4.

D. Real-World Evaluation

The real-world evaluation deploys StereoAdapter-2 on a BlueROV2 in a controlled indoor tank with calibrated geometry and standardized evaluation. The method achieves the reported best real-world performance across cluttered underwater layouts.

  • Hardware Configuration: BlueROV2 experiments use Jetson Orin NX onboard computation, fisheye stereo cameras, and offline rectification to pinhole geometry.Low-level motion control is delegated to an STM32 microcontroller.
  • Scene Setup and Data Collection: The trials use five tank layouts and three teleoperated routes per layout, yielding 15 time-aligned binocular recordings.Glass containers and irregularly shaped stones represent different clutter complexities.
  • Ground-Truth Acquisition: Per-pixel depth references are obtained by aligning AprilTag-based camera poses with a calibrated 3D tank model.Regions without valid surface intersections are excluded from subsequent evaluation.
  • Evaluation Protocol: All methods receive identical rectified image pairs and preprocessing, with absolute depth recovered from disparity using known stereo geometry.Statistics are computed over valid pixels and aggregated across all recordings.
  • Results: StereoAdapter-2 reaches REL 0.1023, RMSE 1.7164, and A1 accuracy 92.56% on BlueROV2 evaluation.The method is reported to show consistent gains in precision and stability across underwater obstacle arrangements.

E. Ablation Study

The ablations examine StereoAdapter-2's architectural components, training hyperparameters, and ConvSS2D design choices. Results identify accuracy–efficiency trade-offs in state dimension and favor cross-scanning for performance.

  • Model Components: The model ablation evaluates the Depth Anything 3 encoder, monocular disparity initialization, context encoder, and update module.
  • Training Hyperparameters: The training ablation studies the effects of different hyperparameter settings.
  • ConvSS2D Hyperparameters: dstate = 16 achieves the best REL and RMSE scores, but increasing state dimension raises computational overhead and reduces throughput.
  • ConvSS2D Hyperparameters: Increasing the SSM expansion ratio beyond 1.0 significantly degrades performance, while dstate = 4 with an SSM ratio of 1.0 is selected for competitive accuracy and high throughput.
  • SS2D Scanning Patterns: Cross-scanning consistently outperforms unidirectional and bidirectional scanning while maintaining the number of scanning directions.

VI. TEST-TIME EFFICIENCY

StereoAdapter-2 is evaluated for end-to-end inference efficiency and deployment behavior alongside its reported accuracy. It achieves the lowest reported latency on Jetson Orin NX, while continuous temporal consistency remains an open limitation.

  • Evaluation Setup: The efficiency evaluation reports per-frame end-to-end latency in milliseconds on Jetson Orin NX 32GB using TensorRT, batch size 1, and 640×320 input resolution.
  • Limitations: The synthetic-to-real gap persists under severe turbidity, strong backscatter, or rapidly varying illumination when training diversity may not capture real-world complexity.
  • Limitations: Temporal consistency during continuous deployment remains challenging because consecutive depth predictions may flicker or become unstable.
  • Method Context: StereoAdapter-2 combines ConvSS2D-based long-range propagation, UW-StereoDepth-80K, and dynamic LoRA adaptation for zero-shot underwater stereo depth estimation.

APPENDIX

Qualitative evaluations compare zero-shot stereo depth predictions across underwater benchmarks and a robotic platform. The reported results emphasize coherent structure, preserved details, and stable real-world predictions, while the dataset samples illustrate varied underwater conditions.

  • SQUID: On SQUID, StereoAdapter-2 produces more coherent disparity maps, clearer object boundaries, and fewer artifacts in textureless and low-contrast regions.
  • Robot Platform: On the robot platform, the proposed method demonstrates stable and consistent depth predictions under real underwater conditions.
  • TartanAir Ocean: On TartanAir Ocean, the method preserves fine-grained structural details and large-disparity regions more effectively.
  • UW-StereoDepth-80K: UW-StereoDepth-80K samples cover diverse underwater scenes and baselines, providing supervision for large-scale underwater stereo adaptation.
Loading 2602.16915v1…