A computer vision end-to-end method for predicting sheep body weight from depth images
Small Ruminant Research · 2026-08 · 261 · 107799
DOI: 10.1016/j.smallrumres.2026.107799 ↗Kurz gefasst
Fachartikel in der Zeitschrift Small Ruminant Research, 2026-08. Verfasst von Ewerton Costa Sousa und 2 weiteren Autor:innen. Die Arbeit nennt 39 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY 4.0). Gefördert von: National Council for Scientific and Technological Development.
Abstract (Originalsprache)
Regular monitoring of animal weight is essential in livestock management; however, traditional weighing methods using scales are labor-intensive and stressful for both animals and humans. To address these challenges, computer vision (CV) techniques have been explored for predicting animal weight from images. However, current works with sheep still suffer from low accuracy and generalization. Within the scope of a coarse-to-fine research approach aimed at developing more efficient, non-invasive, and generalizable methods for small ruminants weighing in livestock farms, this study evaluates the accuracy and generalization of predicting the body weight of sheep from dorsal depth images. It was proposed a ConvNet method based on the EfficientNet-B3 network which operates in an end-to-end fashion: it automatically locates the animal, measures the traits most relevant to prediction, and estimates sheep’s body weight. Model performance was evaluated using the Coefficient of Determination ( R 2 ), Mean Absolute Error (MAE), and Mean Absolute Percentage Error (MAPE). Saliency maps were generated to qualitatively assess feature relevance and model interpretability, and the impacts of combining images from different farms on prediction performance were also analyzed. Moreover, the factors that influenced the performance (experimental, animal, and image-related sources of variation) were investigated using a novel approach based on Generalized Linear Mixed Model (GLMM). The proposed model achieved promising results, with an R 2 of 0.97, MAE of 1.68 kg, and MAPE of 6.74%, demonstrating competitive performance with existing approaches. These findings contributes to advance technological methods for small ruminants meat sector.
Förderung
- National Council for Scientific and Technological Development
Zitieren
APA
Sousa, E. C., Rabelo, R. D. A. L., & Sarmento, J. L. R. (2026). A computer vision end-to-end method for predicting sheep body weight from depth images. Small Ruminant Research, 261, 107799. https://doi.org/10.1016/j.smallrumres.2026.107799
Vancouver
Sousa EC, Rabelo RDAL, Sarmento JLR. A computer vision end-to-end method for predicting sheep body weight from depth images. Small Ruminant Research. 2026;261:107799. doi:10.1016/j.smallrumres.2026.107799
BibTeX
@article{Sousa2026,
author = {Sousa, Ewerton Costa and Rabelo, Ricardo de Andrade Lira and Sarmento, José Lindenberg Rocha},
title = {A computer vision end-to-end method for predicting sheep body weight from depth images},
journal = {Small Ruminant Research},
year = {2026},
volume = {261},
pages = {107799},
doi = {10.1016/j.smallrumres.2026.107799}
}
EUVIMED