Answers are grounded in this paper and cite their sources.
Accurate underwater stereo 3D reconstruction remains challenging for sealed binocular cameras, because refraction through flat viewing windows violates the central projection and linear epipolar geometry assumptions underlying conventional calibration and rectification methods. This paper presents a refraction-aware calibration and rectification framework for underwater stereo 3D reconstruction. The framework adopts an explicit multilayer refraction model to represent the non-central imaging geometry caused by “air-glass-water” refraction. Based on this model, a two-stage air-water calibration strategy is developed, in which the intrinsic and stereo extrinsic parameters are calibrated in air, while the refraction-related parameters, including air-layer thicknesses and glass normal directions, are estimated underwater using multi-objective physical constraints. To enable accurate stereo matching, a depth-range-constrained rectification method is proposed to remap nonlinear refractive correspondence curves into approximately horizontal correspondences within the working range. The resulting disparities are then converted into metric underwater point clouds through refractive ray back-projection and closest-ray triangulation. Experiments were conducted using a self-developed sealed underwater binocular sensor on a water-tank platform. Standard ball-bar measurements and comparisons with structured-light reference scans were performed to assess reconstruction accuracy. Within the 400–600 mm working range, the standard ball-bar experiment achieved an average absolute length error of 0.9580 mm, while comparisons with structured-light reference scans yielded RMSE values of 0.8345 mm and 0.6022 mm for two anchor-chain components. These results demonstrate that the proposed framework provides an accurate geometric foundation for millimeter-level underwater stereo 3D reconstruction.
No takes yet. Share an insight, caveat, or question.
Zhang et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: