Evaluating the robustness of 3-D space-carving seed reconstruction under image noise

Abstract

Three-dimensional (3-D) seed reconstruction is a critical component of automated plant phenotyping, enabling accurate measurement of seed volume and surface morphology from multi-view image data. Among various reconstruction approaches, space carving provides a simple yet powerful voxel-based method that integrates multiple binary silhouettes to estimate the true object geometry. However, the accuracy of such reconstructions depends strongly on the quality and consistency of the image masks used as input. In practical imaging pipelines, image errors arise unavoidably due to lighting variation, sensor artifacts, and imperfect thresholding or learning-based models. This study investigates the robustness of 3-D space-carving reconstruction under con trolled iamge perturbations. A single wheat seed was imaged from 36 viewpoints at 10◦ intervals, and reconstruction was performed using the first 10 silhouettes—sufficient for con vergence of the carved volume—via the multi-threaded carving algorithm of Nielsen1. Binary masks were systematically modified using three perturbation types: morphological erosion, morphological dilation, and random salt-and-pepper noise. A 3 × 3 kernel was applied, and the percentage of pixels altered (0.2–1.0%) was varied to ensure consistent noise scaling across all perturbation modes. The results show that space carving is highly stable under both erosion and dilation, with minor deviations even at the highest perturbation levels, confirming robustness to moderate under- and over-segmentation. In contrast, salt-and-pepper noise caused large deviations in surface area (up to 150%) and moderate reductions in volume (about 13%) due to the formation of small holes and fragmented silhouettes. These findings reveal that while the method is resilient to smooth boundary variation, it is highly sensitive to loss of silhouette connectivity.