3D Plant Phenotyping: All You Need is Labelled Point Cloud Data
Résumé
In the realm of modern digital phenotyping technological
advancements, the demand of annotated datasets is increasing for either
training machine learning algorithms or evaluating 3D phenotyping
systems. While a few 2D datasets have been proposed in the community
in last few years, very little attention has been paid to the construction
of annotated 3D point cloud datasets. There are several challenges
associated with the creation of such annotated datasets. Acquiring
the data requires instruments having good precision and accuracy
levels. Reconstruction of full 3D model from multiple views is a challenging
task considering plant architecture complexity and plasticity, as
well as occlusion and missing data problems. In addition, manual annotation
of the data is a cumbersome task that cannot easily be automated.
In this context, the design of synthetic datasets can play an
important role. In this paper, we propose an idea of automatic generation
of synthetic point cloud data using virtual plant models. Our approach
leverages the strength of the classical procedural approach (like
L-systems) to generate the virtual models of plants, and then perform
point sampling on the surface of the models. By applying stochasticity
in the procedural model, we are able to generate large number of diverse
plant models and the corresponding point cloud data in a fully automatic
manner. The goal of this paper is to present a general strategy
to generate annotated 3D point cloud datasets from virtual models. The
code (along with some generated point cloud models) are available at:
https://gitlab.inria.fr/mosaic/publications/lpy2pc.
Origine : Fichiers produits par l'(les) auteur(s)