💻 computer science

UDR-Mamba: A Depth-Structured Reliability-Guided Mamba Framework for RGB-D Underwater Salient Object Detection

The paper proposes UDR-Mamba, a depth-structured and reliability-guided Mamba framework that transforms raw depth maps into structural cues and selectively fuses them with RGB features via a shared encoder to achieve state-of-the-art performance in underwater salient object detection despite noisy depth measurements.

Original authors: Ning Li, Jiaxin Li, Xiwen Yang, Jianhang Shi, Haojie Wu, Gaochao Yang

Published 2026-09-30

📖 5 min read🧠 Deep dive

CC BY 4.0

Original authors: Ning Li, Jiaxin Li, Xiwen Yang, Jianhang Shi, Haojie Wu, Gaochao Yang

Original paper licensed under CC BY 4.0 (https://creativecommons.org/licenses/by/4.0/). ✨ This is an AI-generated explanation of the paper below. It is not written or endorsed by the authors. For technical accuracy, refer to the original paper. Read full disclaimer

Deep beneath the ocean's surface, where sunlight fades into a hazy blue and colors wash out, robots and cameras struggle to see what is in front of them. This is the world of underwater vision, a field critical for guiding autonomous vehicles, monitoring marine life, and helping humans explore the deep. In these murky environments, standard cameras often fail because the water scatters light, blurs edges, and distorts colors, making it difficult to tell a fish from a rock or a ship from the sea floor. To solve this, scientists have long tried to combine the familiar color images from standard cameras with depth maps, which act like a 3D sketch showing how far away objects are. However, underwater depth maps are often noisy and unreliable, filled with gaps and errors that can confuse a computer just as much as the murky water confuses a human eye. The challenge has been to teach a machine to use the helpful parts of these 3D sketches while ignoring the parts that are broken or misleading.

A team of researchers at Changzhou University has developed a new approach to this problem, creating a system they call UDR-Mamba. Instead of simply feeding raw color and raw depth data into a computer, their method first takes the messy depth information and reorganizes it into a clearer, more useful format. Imagine taking a blurry, grainy photograph and then tracing the outlines of the objects within it while also highlighting the areas where the texture changes sharply; this is essentially what their system does with depth data. They break the depth map down into three distinct components: a normalized version that shows the overall layout, a map of the edges where depth changes suddenly, and a measure of local contrast that highlights small geometric differences. By presenting the computer with these three structured views of the depth, the system can better understand the shape and boundaries of objects, even when the original depth data is imperfect.

The core of their innovation lies in how the system decides which parts of the depth information to trust. In many underwater scenes, the depth sensors produce data that is simply wrong in certain spots, perhaps due to bubbles or particles in the water. The researchers designed a mechanism that acts as a gatekeeper, constantly evaluating the reliability of the depth information at every stage of processing. It looks at the color image, the structured depth data, and the differences between them to decide how much weight to give the depth cues. If the depth information looks trustworthy, the system blends it in to sharpen the image; if it looks noisy or unreliable, the system ignores it to prevent errors. This selective process ensures that the computer only uses the depth data when it is helpful, effectively filtering out the noise that usually plagues underwater vision.

To tie these components together, the researchers employed a modern type of artificial intelligence architecture known as a state-space model, specifically a variant called Mamba. Unlike older systems that process images in small, isolated chunks, this architecture is designed to understand long-range connections across the entire image, allowing it to see the big picture of a scene while still noticing fine details. The team used a single, shared set of brain-like parameters to process both the color and the depth data, which makes the system more efficient and less redundant than previous methods that used separate, heavy processors for each type of information. This shared approach allows the system to learn the relationship between color and depth without needing a massive amount of computing power, striking a balance between speed and accuracy.

When tested on large collections of underwater images, the new system demonstrated a clear ability to find and outline important objects more accurately than existing methods. On a standard benchmark containing over ten thousand underwater images, the system achieved a high level of precision in identifying the shape and location of salient objects, outperforming previous state-of-the-art techniques in several key measures of accuracy. It was particularly effective at handling scenes with low contrast and complex backgrounds, where other systems often produced incomplete outlines or missed objects entirely. The researchers found that by explicitly modeling the structure of the depth data and gating its reliability, the system could recover sharp boundaries even in the most challenging conditions.

The results suggest that the key to better underwater vision is not just having more data, but having smarter ways to interpret the data that is available. By transforming raw depth maps into structured cues and carefully filtering them through a reliability check, the system overcomes the inherent limitations of underwater sensors. This work provides a practical path forward for developing more robust robotic vision systems that can operate effectively in the difficult, light-starved environments of the ocean, potentially improving everything from environmental monitoring to underwater exploration. The researchers note that while their current system is highly effective, future work will focus on making it even lighter and more robust, ensuring it can run on the limited hardware of real-world underwater robots.

Technical Summary: UDR-Mamba

Problem Statement
Underwater salient object detection (USOD) is critical for robotic navigation, marine perception, and ecological monitoring. However, accurate segmentation is severely hindered by the unique degradation of underwater imagery, including color shifts, illumination attenuation, scattering, low contrast, and blurred boundaries. These factors diminish the discriminative power of standard RGB appearance features. Furthermore, while depth maps offer complementary geometric cues, raw underwater depth data is often unreliable, containing noise, missing regions, and inaccurate boundaries. Existing RGB-D methods often fuse depth indiscriminately or treat it as a simple single-channel auxiliary input, failing to explicitly model depth structural cues (such as edges and local contrast) or to adaptively assess the reliability of depth information during cross-modal fusion. Additionally, while Vision Transformers and Mamba models offer strong global context modeling, simply replacing backbones does not resolve the specific challenges of uncertain depth quality in underwater scenes.

Methodology: UDR-Mamba
The authors propose UDR-Mamba, a depth-structured and reliability-guided framework built upon a shared-weight VMamba backbone. The architecture consists of four primary components:

Depth Structuralization Input Module (DSI): Instead of feeding raw depth maps directly, the DSI transforms a single-channel depth map into a three-channel structural representation:
- Normalized Depth ( $D_n$ ): Captures overall spatial layout.
- Depth Edges ( $E$ ): Extracted via Sobel operators to highlight structural discontinuities and object contours.
- Local Depth Contrast ( $C$ ): Computed as the absolute difference between local depth and the local mean (via 7×7 pooling) to enhance local geometric variations.
  This structured input explicitly provides the network with boundary and local geometric cues.
Shared-Weight VMamba Encoder: The framework employs a two-stream encoding strategy where RGB images and the structured depth maps pass through separate forward passes but share the same VMamba backbone parameters. This design reduces parameter redundancy while maintaining the representational capacity of a dual-stream architecture. The encoder outputs four levels of multiscale features.
Depth-Reliable Mamba Fusion Module (DRMF): To address the issue of noisy depth data, the DRMF performs layer-wise adaptive fusion. It estimates a reliability gate ( $G_i$ ) based on the current RGB features, aligned depth features, cross-modal discrepancies, and resized depth-edge cues. This gate selectively injects trustworthy depth information into the RGB representation via element-wise multiplication, preventing unreliable depth noise from degrading the prediction. The module also incorporates Compact Mamba Context to enhance long-range dependency modeling within the fused features.
Mamba Pyramid Decoder: A single-head decoder aggregates multiscale fused features in a top-down manner. It utilizes Compact Mamba Context for global context enhancement and introduces depth-edge information at the final stage for detail refinement. The model is trained using only the final saliency prediction supervised by Binary Cross-Entropy (BCE) and soft Intersection-over-Union (IoU) losses, avoiding complex auxiliary loss functions.

Key Contributions

Two-Stream VMamba Framework: The authors introduce a shared-weight VMamba backbone for RGB-D USOD, leveraging visual state-space modeling to enhance long-range contextual reasoning in complex underwater environments.
Depth Structuralization: The DSI module explicitly transforms raw depth into normalized depth, edges, and local contrast, fully exploiting geometric and boundary cues rather than treating depth as a raw auxiliary channel.
Reliability-Gated Fusion: The DRMF module adaptively estimates depth reliability from cross-modal discrepancies and edge information, enabling the selective injection of trustworthy depth data while suppressing noise.
Efficiency and Simplicity: By sharing encoder weights and using a single-head decoder with a simple loss objective, the framework achieves a favorable trade-off between accuracy and computational efficiency.

Experimental Results
The method was evaluated on the USOD10K and USOD datasets against state-of-the-art methods (e.g., HDANet, Dual-SAM, SPDE).

USOD10K: UDR-Mamba achieved the highest scores in $S_m$ (0.9332), $E_{max}$ (0.9713), and $F_{max}$ (0.9365), with an MAE of 0.0192. It outperformed the closest competitor (HDANet) in $S_m$ and $F_{max}$ .
USOD: The method achieved the highest $E_{max}$ (0.9573) and the lowest MAE (0.0359), with the second-highest $F_{max}$ (0.9314).
Ablation Studies:
- DSI: Replacing the structured input with raw depth reduced performance across all metrics (e.g., $S_m$ dropped from 0.9332 to 0.9236), confirming the value of explicit edge and contrast modeling.
- DRMF: Removing the reliability gate resulted in lower performance, demonstrating that indiscriminate fusion degrades accuracy.
- Weight Sharing: The shared-weight configuration reduced parameters by approximately 49.91M compared to independent dual backbones while maintaining nearly identical accuracy, validating the efficiency of the design.

Significance and Claims
The paper claims that UDR-Mamba demonstrates that explicit depth-structure modeling and reliability-guided fusion significantly improve boundary-aware multimodal representations in underwater scenes. The authors assert that their approach effectively addresses the dual challenges of degraded RGB appearance and unreliable depth measurements without relying on complex auxiliary losses or heavy computational overhead. The results suggest that the proposed framework provides a robust and efficient solution for RGB-D underwater salient object detection, offering superior structural consistency and boundary localization compared to existing methods. Future work is noted to focus on lightweight Mamba architectures for real-time deployment and improved depth-quality assessment mechanisms.

Technical Summary: UDR-Mamba

More like this