Fig 1.
Sample data of a maize (A) and a tomato plant (B) scanned periodically.
Temporally consistent labels are assigned to each individual leaf, as indicated by color.
Table 1.
Our dataset contains a considerably higher number of point clouds with respect to other public datasets.
Note that the ETH dataset is meant to develop algorithms to evaluate plant stress.
Fig 2.
Tomato plants (A) and maize plants (B) in pots in the growing station and the process of measuring with the scanner of a tomato plant (C).
Fig 3.
Scanning device (A), complete scanning system (B) consisting of the scanning device attached to the measuring arm, and the scanning system in the measurement laboratory environment (C).
Fig 4.
Raw point clouds of a tomato plant (A) and a maize plant (B).
For illustration purpose the point size has been increased in the visualization.
Fig 5.
Detailed view of a tomato plant point cloud (A) and its segmentation into stem and individual leaves, illustrated in different colors (B).
Fig 6.
Point cloud of a maize plant (A), segmentations derived according to the Leaf Collar Method (B) and according to the Leaf Tip Method (C).
Fig 7.
Labeling a maize plant into stem and leaves.
Detailed view of a maize point cloud (A) and its segmentation derived according to the Leaf Collar Method (B). The same maize plant two days later (C) and its segmentation derived according to the Leaf Collar Method (D).
Fig 8.
Dataset coverage of the 7 maize plants (top) and the 7 tomato plants (bottom) with respect to the day within the measurement period ( = labeled point clouds available,
= only point clouds available,
= no data available).
Fig 9.
Semantic segmentation comparison between different deep learning approaches (Ground-truth (A), LatticeNet (B), PointNet++ (C) and PointNet (D)).
PointNet++ misses parts of the stem and PointNet misses it completely. In contrast, LatticeNet segments the whole stem accurately.
Table 2.
Mean Intersection-over-Union and IoU per class results for three different neural network architectures tested on our maize and tomato point clouds.
Table 3.
Segmentation performance of different deep learning on Pheno4D dataset.
Fig 10.
Time series of a tomato plant scanned in various days together with the extracted skeleton.
Fig 11.
Individual organs are segmented and skeletons are fitted (A). The skeletons corresponding to different scanning days are associated temporally. The correspondences allows to non-rigidly deform one cloud in order to match the other (B).
Fig 12.
Raw point cloud (A), initial mesh extracted using Poisson reconstruction (B) and mesh after trimming of triangles with low density (C).
Fig 13.
Tracking of phenotyping traits starting from raw point clouds.