One of the biggest challenges of deep neural network to perform segmentation of point clouds is the requirement of large amount of annotated data, which is expensive in terms of manual labour and time. For the case of plants, there is a scarcity of datasets to train the networks to perform organ segmentation for automated phenotyping applications.
In this work, we explore how the use of virtual plants as modelled by stochastic L-systems can circumvent this problem. We investigate the effect of point density and how the complexity of the plant model affect the transfer of learning from virtual to real plants on a segmentation task based on PointNet++.
