A Simple Approach to Pavement Cell Segmentation

Rostislav Shepel, Sotirios A. Tsaftaris, Mario Valerio Giuffrida

CVPPA 2024 Workshop at ECCV

Shepel, R., Romanowski, A., Giuffrida, M.V. (2025). A Simple Approach to Pavement Cell Segmentation. In: Del Bue, A., Canton, C., Pont-Tuset, J., Tommasi, T. (eds) Computer Vision – ECCV 2024 Workshops. ECCV 2024. Lecture Notes in Computer Science, vol 15625. Springer, Cham. https://doi.org/10.1007/978-3-031-91835-3_16

wp-content/uploads/2020/10/tex.png
Get Paper
Code on GitHub
@InProceedings{10.1007/978-3-031-91835-3_16,
author="Shepel, Rostislav
and Romanowski, Andrew
and Giuffrida, Mario Valerio",
editor="Del Bue, Alessio
and Canton, Cristian
and Pont-Tuset, Jordi
and Tommasi, Tatiana",
title="A Simple Approach to Pavement Cell Segmentation",
booktitle="Computer Vision -- ECCV 2024 Workshops",
year="2025",
publisher="Springer Nature Switzerland",
address="Cham",
pages="240--251",
abstract="This study focuses on segmenting pavement cells from microscopy images of Arabidopsis thaliana plants, which is critical for linking cellular traits to overall plant performance. Differently than the current state-of-the-art, we propose a simple, easy-to-train approach using partially annotated datasets to address the challenges of irregular pavement cell shapes. Specifically, we employed U-Net and DeepLabV3 architectures for segmentation, showing that both models can perform well despite the constraints. Post-segmentation, we used PaCeQuant to extract phenotyping data, demonstrating the effectiveness of our method. The results indicate that U-Net provides a slightly closer match to the true mask, though DeepLabV3 also performs robustly. This approach facilitates more accurate and efficient plant phenotyping, contributing to sustainable agricultural practices. Code is publicly available at the following repository: https://github.com/Rosti35/pavement-cell-segmentation.",
isbn="978-3-031-91835-3"
}

Abstract

This study focuses on segmenting pavement cells from microscopy images of Arabidopsis thaliana plants, which is critical for linking cellular traits to overall plant performance. Differently than the current state-of-the-art, we propose a simple, easy-to-train approach using partially annotated datasets to address the challenges of irregular pavement cell shapes. Specifically, we employed U-Net and DeepLabV3 architectures for segmentation, showing that both models can perform well despite the constraints. Post-segmentation, we used PaCeQuant to extract phenotyping data, demonstrating the effectiveness of our method. The results indicate that U-Net provides a slightly closer match to the true mask, though DeepLabV3 also performs robustly. This approach facilitates more accurate and efficient plant phenotyping, contributing to sustainable agricultural practices. Code is publicly available at the following repository: https://github.com/Rosti35/pavement-cell-segmentation.