Hierarchical Reconstruction Module for Multi-View 3D Reconstruction of Fruit Trees

This paper proposes a novel Hierarchical Reconstruction Module (HRM) for multi-view 3D reconstruction of fruit trees. The HRM can effectively generate images of missing viewpoints using limited input views and integrates seamlessly with the MVSTER model. Through a hierarchical decomposition strategy and core point localization mechanism, HRM achieves 86.3% accuracy in branch topology reconstruction while maintaining a lightweight parameter size of only 159KB. We also introduce a Combined Annealing with Decay (CA WD) learning rate adjustment strategy that significantly enhances model performance. This work provides an efficient solution for practical orchard management applications by reducing data acquisition requirements while maintaining reconstruction accuracy.