Balancing accuracy, completeness, and efficiency for rice 3D reconstruction through CBAM-UNet-based multi-view segmentation and camera configuration optimization.

Multi-view 3D reconstruction has been widely applied in plant phenotyping, but the complex canopy structure of rice plants poses significant challenges for reconstruction accuracy, completene7ss, and efficiency. In this study, we developed an optimized workflow for 3D reconstruction of rice using the self-developed Metatlas V1 multi-view imaging platform combined with the COLMAP + OpenMVS pipeline. A CBAM-UNet-based image segmentation model was used to extract plant regions from complex backgrounds, outperforming thresholding, U-Net, and U-Net++, and increasing the number of reconstructed points. Camera configuration optimization was performed by systematically combining equidistant and greedy strategies, resulting in an optimal six-camera setup (#2, #3, #4, #6, #7, and #8, corresponding to -30°, -15°, 0°, +30°, +45°, and +60° relative to the horizontal viewpoint), which provided complementary views of the canopy inner structure and stem base. This configuration reduced the average nearest-neighbor distance to 0.037-0.066 cm across different varieties and growth stages compared with the full 11-camera setup, retained 73-76% of the points, and decreased reconstruction duration by approximately 70-80%. The strong agreement between point-cloud-derived and manually measured plant height and canopy width (