Self-Supervised Learning With RGB-Guided Attention and Individual Consistency for Pig Weight Estimation
Smart Agricultural Technology · 2026-11 · 102717
DOI: 10.1016/j.atech.2026.102717 ↗Kurz gefasst
Fachartikel in der Zeitschrift Smart Agricultural Technology, 2026-11. Verfasst von Sizhuo Li und 4 weiteren Autor:innen. Die Arbeit nennt 34 Literaturangaben. Der Volltext ist frei zugänglich (Open Access, CC BY-NC-ND 4.0).
Abstract (Originalsprache)
Monitoring body weight during pig growth is critical for precision livestock management. Non-contact weight estimation techniques can reduce stress responses in pigs and improve operational efficiency. Many multimodal methods for pig weight estimation rely on supervised pre-trained backbones, while self-supervised learning remains underexplored in this task. Furthermore, fusion mechanisms in these methods often adopt symmetric architectures without using the semantically stronger modality to guide the other. To address these limitations, a method named DINOv3-DFR (DINOv3-based Dual-stream Fusion Regressor) is proposed, built upon self-supervised pre-training. Specifically, the proposed method designs an RGB-guided attention module, where the pig appearance cues learned from RGB stream serve as queries to selectively enhance the pig spatial structure cues learned from depth stream by leveraging gated fusion during the model’s feature learning stage. Based on the fused features, a multi-level feature aggregator reconstructs high-resolution representations for preserving both global body shape and local contour details of live pigs. Additionally, many existing image-level loss functions focus only on single-image accuracy, ignoring that multiple same-pig images within a short time period should yield consistent weight predictions. Therefore, an Individual-level Consistency Loss is introduced to enforce the inherent weight stability of each pig during training. Experiments are conducted on a self-constructed top-view RGB-D dataset of pigs with a training set and a testing set containing 8,162 and 4,193 RGB-D images, respectively. The experimental results demonstrate that the proposed DINOv3-DFR achieves a mean absolute error of 2.96 kg and a mean absolute percentage error of 3.47%, and an R² score of 0.927, significantly outperforming existing approaches.
Zitieren
APA
Li, S., Mi, Y., Huang, H., Ding, X., & Li, T. (2026). Self-Supervised Learning With RGB-Guided Attention and Individual Consistency for Pig Weight Estimation. Smart Agricultural Technology, 102717. https://doi.org/10.1016/j.atech.2026.102717
Vancouver
Li S, Mi Y, Huang H, Ding X, Li T. Self-Supervised Learning With RGB-Guided Attention and Individual Consistency for Pig Weight Estimation. Smart Agricultural Technology. 2026:102717. doi:10.1016/j.atech.2026.102717
BibTeX
@article{Li2026,
author = {Li, Sizhuo and Mi, Yang and Huang, Hai and Ding, Xiangdong and Li, Tao},
title = {Self-Supervised Learning With RGB-Guided Attention and Individual Consistency for Pig Weight Estimation},
journal = {Smart Agricultural Technology},
year = {2026},
pages = {102717},
doi = {10.1016/j.atech.2026.102717}
}
EUVIMED