Geometric-constraint-aware binocular framework for robust absolute phase recovery in structured light 3D reconstruction

To address the critical challenges in balancing reconstruction accuracy, computational efficiency, and robustness to complex surfaces faced by optical three-dimensional reconstruction systems using structured light projection, this paper presents a novel binocular deep learning framework, to the best of our knowledge, that synergistically integrates geometric constraints with data-driven learning for robust absolute phase recovery. Our key contributions include: (1) a dual-view structured light system with a StereoPhase benchmark dataset containing 23,000 synchronized binocular phase maps of both virtual and real-world ground-truth 3D models, specifically for challenging scenarios like discontinuous surfaces; (2) a dual-path network architecture GCANet featuring cross-view feature fusion modules and dilated residual blocks to address the phase jumps; and (3) a consistency-aware loss function that applies epipolar geometric constraints of two perspectives. Extensive experiments demonstrate competitive performance between speed and accuracy, achieving lower error in all our datasets compared to existing learning-based approaches while maintaining real-time processing. The framework establishes what we believe to be a new paradigm for phase unwrapping that effectively bridges the gap between traditional geometric methods and modern deep learning solutions, particularly benefiting industrial inspection and biomedical applications requiring both precision and computational efficiency.