Development of a Framework for Self-Supervised Segmentation of Image Time Series Using Optical Flow

Abstract

In the field of plant biology, understanding root growth dynamics is crucial for various applications, ranging from crop genetic improvement to the study of physiological responses under different environmental conditions. To analyze this dynamic behavior through time-lapse image sequences or videos, it is essential to segment the roots in order to ultimately phenotype them, thereby extracting parameters such as length, growth angle, among others. Currently, the most commonly used segmentation methods are based on convolutional neural networks (CNNs), which require a large amount of manually labeled data for training. This work aims to address this challenge through the development of a self-supervised method for label generation, leveraging the temporal redundancy inherent in videos of growing roots. This approach significantly reduces the need for manual annotations, making the segmentation process more efficient.