A 3D Light weight neural network for plant part segmentation and
architectural trait extraction
Plant architecture is an important contributing factor for enhanced
yield production and quality. The architecture traits are analyzed for
crop health monitoring and genetic manipulation for generating high
yielding varieties. Computer vision methods applied on 3D pointcloud
allow more accurate extraction of architecture traits but consume more
time and memory compared to 2D images. This study aims to design light
weight 3D deep network for Cotton plant part segmentation and derive
seven architectural traits of mainstem height, mainstem diameter, branch
inclination angle, branch diameter, and number of branches, nodes, and
cotton bolls. The pointcloud data is collected using FARO LiDAR scanner.
The mainstem, branches and cotton bolls are manually annotated using
Open3D. The preprocessing steps of denoising, normalization and down
sampling are applied. 3D Deep network is designed to sample 1024, 512
and 256 points where neighborhood aggregation is performed at radius
levels of 1cm, 5cm, and 30cm respectively. Features for remaining points
are interpolated. The features from each radius level are concatenated
and passed to multi-layer perceptron for pointwise classification.
Results indicate that mean IoU and accuracy of 84% and 94% are
achieved respectively. A 6.5 times speedup in inference time and 2.4
times reduction in memory consumption compared to Pointnet++ is gained.
After applying postprocessing on part segments, an R square value of
more than 0.8 and mean absolute percentage error of less than 11% are
achieved on all derived architecture traits. The trait extraction
results indicate potential utility of this process in plant physiology
and breeding programs.