Reconstruction with Underwater Stereo Vision

Lin Gui, Zhiqiang Wei, Bo Yin, Lei HuangAdvances 2014 · DOI

Abstract

Conventional multi-view theory fails to explain underwater stereo vision because refraction bends light rays. This paper focuses on one of the most common scenarios of underwater vision: a stereo vision system of two cameras with a flat housing. In this setup, light rays refract twice—once at the air-housing interface and once at the housing-water interface. The paper models the geometry of light propagation explicitly. Under the principle of light ray reversibility, the discussion follows a "backward projection" approach, casting a light ray from the camera center to the object point. After proving the coplanarity of the entire light ray path, the 2D case is extended to 3D space. A concise and feasible 3D reconstruction algorithm is proposed, requiring no specialized configuration or additional hardware. Experimental results demonstrate high accuracy and a low error rate.

Introduction: The Challenge of Refraction

Standard computer vision algorithms typically rely on the pinhole camera model, which assumes that light travels in straight lines from the object to the sensor. However, this assumption breaks down in underwater environments. When a camera is placed inside a flat glass housing, light rays must pass through three different media: air, glass, and water.

As light crosses these boundaries, it bends according to Snell’s Law. This double refraction causes significant distortion, making objects appear closer and slightly warped. Mathematically, "forward projection"—mapping a 3D point in water to a 2D pixel on the sensor—becomes a complex, non-linear problem that is difficult to solve analytically. Many existing solutions attempt to simplify this by requiring "axial configuration," where cameras are perfectly perpendicular to the glass. However, such precise alignment is often impossible in field deployments like underwater robotics.

Geometric Modeling via Backward Projection

To overcome these challenges, the authors propose a model based on the principle of light reversibility. Instead of calculating where a 3D point lands on the image, they utilize "backward projection." This involves starting at the camera center and "casting" a ray through a specific pixel, tracing its path out into the water.

Geometry of light propagation through air, glass, and water interfaces.

A key technical contribution of this work is the proof of coplanarity. The authors demonstrate that the entire path of a light ray—from the camera center ($C$), through the air-glass interface ($P_a$), through the glass-water interface ($P_g$), and finally to the object point ($P_w$)—lies within a single 3D plane. This plane is defined by the initial ray direction and the surface normal of the housing. By proving coplanarity, the complex 3D bending problem is reduced to a series of 2D vector calculations, making the 3D reconstruction math tractable and efficient.

The model accounts for the thickness of the glass housing and the refractive indices of all three media. To determine the exact orientation of the glass, a checkerboard marker is used for "housing calibration," allowing the system to compute the surface normal of the glass relative to the camera.

Experimental Validation

The proposed algorithm was tested using a custom experimental setup. The hardware consisted of two Canon 5D cameras positioned to view a 60cm x 25cm x 40cm glass water tank. The tank walls were 0.65cm thick.

The physical experimental setup used for 3D reconstruction validation.

Submerged objects, including a mug and a patterned plate, were used as targets. These objects provided high-contrast features necessary for the SIFT (Scale-Invariant Feature Transform) algorithm, which the authors used to find corresponding points between the two stereo views. By matching these features and applying the backward projection model, the system could reconstruct the 3D coordinates of points on the objects' surfaces.

Results and Discussion

The experimental results confirmed the accuracy and feasibility of the proposed method. The system achieved an average reconstruction error of 4.13 mm at a distance of approximately 0.6 meters. This corresponds to an overall error rate of less than 2%.

Qualitatively, the 3D point clouds successfully recovered the depth and structure of the submerged objects, effectively compensating for the visual distortions caused by the water and glass. The flexibility of the model is a major advantage, as it does not require the cameras to be perfectly aligned with the housing, making it highly practical for Remotely Operated Vehicles (ROVs).

Conclusion and Future Work

While the geometric model is highly effective, the authors identified several areas for improvement. One limitation is the loss of features during the SIFT matching process, often caused by reflections on the glass housing or the unique optical properties of water. Furthermore, the current model focuses solely on the geometry of refraction and does not account for light scattering (turbidity), which can reduce contrast in murkier water.

Future research will focus on developing more robust feature detection methods and incorporating light scattering models to enhance the system's performance in challenging underwater conditions.

Summary updated

Abstract

Conventional multi-view theory fails to explain underwater stereo vision because of the refraction bends light rays. This paper focuses on one of the most common scenarios of underwater vision, a stereo vision system of two cameras with a flat housing, in which the light ray refracts twice, once at the air-housing interface, and once at the housing-water interface.The papermodelsthe geometry of the light propagation explicitly. Under the principle of light ray's reversibility,discussion on one camera's view in 2D follows the backward projection by casting a light ray from the camera center to the object point.After the entire light ray's coplanarity is proved, the 2D case extends to 3D space easily. In 3D case, computation of object point from the casted ray is derived from the stereo view model. An 3D reconstruction algorithm is also proposed. The algorithm is concise and feasible because no particular configuration or additional device is required. It is proved accurate and low error rate in the experiment.
BibTeX
@inproceedings{gui2014reconstruction,
  author = {Lin Gui and Zhiqiang Wei and Bo Yin and Lei Huang},
  title = {{Reconstruction with Underwater Stereo Vision}},
  booktitle = {Advances},
  year = {2014},
  doi = {10.2991/lemcs-14.2014.227},
}
Figure