Multi-Layer Attention Fusion and Edge-Guided Structural Enrichment Completion Network for Leaf Point Cloud

Three-dimensional (3D) reconstruction is an important task to capture the spatial structure and morphological features of plants. However, the occlusion between leaves resulted in an incomplete point cloud. It thus affects the accuracy of plant phenotype analysis. To obtain complete leaf point clouds, a Multi-Layer Attention Fusion and Edge-Guided Structural Enrichment Network (MAEGNet) was proposed for the Leaf Point Cloud Completion. It designed multiple layers of attention fusion modules and an edge-preserving loss function, using a Transformer to process and fuse local and global features of leaves at varying receptive fields. It encouraged the model to recover the leaf edge structure by calculating the loss between the edges extracted from the completed and ground-truth leaf point clouds. Experimental results show that our model can achieve a better result on our constructed leaf point cloud dataset. The Earth Mover's Distance (EMD) is reduced by 7.44 cm.