Breast morphology assessment and bra structure design via binocular stereovision reconstruction

Purpose To improve the convenience and operability of data acquisition in human body reconstruction, this study aims to propose a binocular stereo vision-based three-dimensional (3D) reconstruction framework for rapid and stable breast morphology reconstruction in women, thereby providing reliable data support for subsequent garment pattern generation. Design/methodology/approach A binocular stereo vision camera was employed to capture dual-view images for breast morphology reconstruction. Building upon the iterative optimization framework of Recurrent All-Pairs Field Transforms (RAFT), an improved stereo matching method was developed by integrating a multi-scale feature extraction network, a learnable nonlinear similarity function, and a cross-connected multi-level Gated Recurrent Unit (GRU) architecture to achieve robust coarse-to-fine disparity refinement. The calibrated and rectified image pairs were then used to generate high-resolution depth maps and dense three-dimensional point clouds. Reconstruction accuracy was validated through registration between the reconstructed point clouds and reference data acquired from a 3D body scanner. Findings The average registration distance between the reconstructed point clouds and TC2 scanning results was 2.66 mm, with maximum deviations below 5 mm for all subjects. The proposed method demonstrated stable performance across different breast morphology categories, with an average reconstruction time of less than 30 s. Measurements extracted from the reconstructed models showed high consistency with manual measurements and exhibited better stability and robustness than 3D scanning results. Originality/value The proposed framework achieves accurate and cost-effective breast morphology reconstruction under weak-texture conditions. By directly extracting anthropometric parameters from reconstructed point clouds for conventional pattern-making, it establishes a practical link between three-dimensional reconstruction and bra pattern generation.