Abstract

3D scans of real world objects are often represented by point clouds, creating XYZ-coordinates of individual scan points. However, unlike point clouds that are generated from CAD data, points generated from a real world scene lack information about their local context, making segmentation of the structural information contained in the data difficult. Using neural networks (e.g. PointNet) has shown promising results. However, this approach is not well suited for scans of large areas of similar objects, like e.g. a wheat field, because of limitations of the input vector size of the neural network. In addition, point clouds are often unordered, further complicating processing. Since point clouds of biological objects often contain recurring features, we propose to subdivide the point cloud into locally neighboring subsets with a fixed number of points. The collection of subsets can then be used to train neural networks. This approach preserves the original resolution of the point cloud while offering simple data augmentation concepts like creating a number of different subset collections from the same ground truth. There are several advantages to this approach, like significantly simplifying the training phase, because a single, large annotated scan can be sufficient for training, utilizing the similarity of the instances of a plant in the field.