Plant leaf point cloud completion based on deep learning

Accurate 3D point cloud acquisition of plant leaves has been widely found in the field of vegetation structure modeling, which is further critical in quantitative remote sensing. Owing to the occlusion between the plant leaves and the limited performance of 3D data acquisition sensors, the acquired leaf point cloud may be incomplete. It is necessary to complete the partial leaf point cloud by some means. Existing point cloud completion methods include registration methods, geometry-based methods and database-based methods, which are time consuming and less effective. This paper proposes a method of plant leaf point cloud completion by using deep Encoder-Decoder framework. The encoder reads incomplete plant leaf point cloud into a shape feature vector and the decoder is trained to predict the complete leaf point cloud. The loss function consists of forward loss and backward loss. For further study, a leaf point cloud dataset is established. The data enrichment is performed by random rotation, random occlusion, random transformation of point cloud sequence, so that the dataset is more representative. The experimental results show that the missed leaf point cloud can be well completed. Meanwhile, the proposed method can directly operate on raw point cloud with less computation and is robust to noisy point cloud.