Three-Dimensional Plant Reconstruction with Enhanced Cascade-MVSNet

Three-dimensional reconstruction is an important method for recovering the morphological structure of plants, and a complete and accurate 3D point cloud of plants can better reflect the phenotypic parameters of plants, such as plant height, volume of the plant and leaf area. In order to obtain more complete and accurate plant point clouds, this paper proposes a series of enhancements based on the Cascade-MVSNet network. To improve the reconstruction of weakly textured areas of plants, a lightweight attention mechanism is incorporated into the feature extraction stage. To enhance the valid points of the plant point clouds and suppress invalid points, focal loss is employed as the loss function, treating the depth estimation problem as a classification task to obtain more accurate depth information. Additionally, to improve the efficiency of the model, we substitute standard convolutions with depth-wise separable convolutions, reducing the number of parameters and computational complexity while maintaining performance. The above work is applied to the plant dataset and the reconstructions showed a significant reduction in invalid points and a clearer reconstructed point cloud.In addition to conducting research on the plant dataset, we also evaluate our approach on the publicly available DTU dataset. Experimental results demonstrate a noticeable increase in reconstruction completeness on the DTU dataset, achieving competitive performance overall.