Abstract: Aiming at the prevalent issues of prolonged reconstruction duration and suboptimal reconstruction quality inherent in existing three-dimensional (3D) reconstruction methods for crops, this study proposes and constructs a novel algorithm specifically designed for reconstructing seedling-stage maize plants. This algorithm is based on an improved framework of the Oriented FAST and Rotated BRIEF-SLAM2 (ORB-SLAM2) system. The core methodological innovation involves the integration of several advanced computer vision techniques to achieve a robust and efficient reconstruction pipeline. Initially, the Fast Library for Approximate Nearest Neighbors (FLANN) algorithm is employed in combination with the Random Sample Consensus (RANSAC) algorithm to perform high-accuracy feature matching on multi-view images of maize seedlings. This hybrid matching strategy effectively identifies and aligns corresponding keypoints across different images while filtering out outliers and incorrect matches, thereby establishing a reliable geometric foundation for subsequent 3D processing. Following successful feature matching and sparse point cloud generation through the modified ORB-SLAM2 front-end, a Multi-View Stereo (MVS) algorithm is seamlessly integrated into the pipeline. This MVS module utilizes the precise camera pose estimations and the established feature correspondences to perform dense matching and depth estimation, ultimately generating a dense, detailed, and accurate 3D point cloud model that faithfully represents the intricate morphology and spatial structure of the target maize plants. To rigorously evaluate the practical utility and accuracy of the reconstructed models, this research further utilizes them for the digital extraction of key architectural parameters, such as plant height, leaf length, and stem diameter, from two distinct maize cultivars. These digitally acquired parameters are then systematically compared against traditional, manually measured ground-truth values to quantitatively validate the accuracy, reliability, and overall effectiveness of the proposed non-contact, image-based phenotyping methodology. Comprehensive experimental results demonstrate the superior performance of our integrated approach. Specifically, the FLANN+RANSAC combination achieves a notably high feature matching correctness rate of 89.00%. The dense reconstruction phase consistently produces detailed point clouds with an average density of 7.13×10⁵ points per model. Remarkably, the entire reconstruction process, from image input to dense point cloud output, requires an average time of only 15.32 min, highlighting its significant efficiency advantage. Furthermore, the errors associated with extracting all major architectural parameters from the 3D models for seedling maize plants are consistently maintained within a 10% threshold. Statistical analysis also confirms a strong and significant correlation between the digitally extracted values and the manual measurements, substantiating the methodological robustness. In conclusion, the proposed algorithm successfully addresses the dual challenges of speed and precision in plant phenotyping. It is characterized by substantially reduced computational time without compromising reconstruction quality, offering a fast, accurate, and non-destructive solution. Therefore, this work provides a solid theoretical foundation and a practical, high-performance technical framework for the automated acquisition of 3D architectural traits in maize plants during the critical seedling stage. This capability holds considerable potential for advancing high-throughput plant phenotyping, informing precision agriculture management decisions, and accelerating crop genetics and breeding research by enabling efficient, objective, and detailed morphological analysis.