This dissertation aims to advance current segmentation algorithms in computer vision with
applications in plant phenotyping. Current segmentation techniques for plants have been
inefficient in computation and data usage. I leverage multi-dimensional (perspective, modality,
temporality) imagery from high-throughput facilities to (1) expose the bias in algorithm
performance in current cosegmentation algorithms; (2) introduce a novel, unsupervised, endto-
end trainable cosegmentation deep learning-based algorithm to segment dynamic, evolving
objects (plants) accurately; and (3) design a new class of phenotypes to aid the understanding
of dynamic plant growth in its environment. These contributions will benefit the future
direction of computer vision to handle more dynamic, evolving objects and guide the process
of cropping systems through imagery.