Paul Nelson Becker · Elias Marks · Yue Linn Chong · Theodor Kapler · Jens Behley · Louis Bonerath · Cyrill Stachniss
[Click to expand]
Robots that can monitor and manage fields have the potential to change agriculture. By allowing individual plant treatment, they can enable growing mixed varieties on agricultural fields, increasing resilience, biodiversity, and soil quality. We study few-shot plant variety classification using self-supervised representations to reduce the need for labeled data. Using zero-shot segmentation, we adapt multi-view self-supervised learning by positioning image crops on leaves, based on the observation that discriminative features between plant varieties are often concentrated on their leaves. We introduce a new image dataset containing 98 sugar beet varieties enabling the study of variety classification. On unseen varieties, our method achieves high accuracies, outperforming existing self-supervised and foundation-model representations and substantially narrowing the gap to supervised representations. Our results also suggest that for self-supervised representations for few-shot fine-grained recognition, where views are placed can be more important than how views are augmented.git clone [email protected]:PRBonn/LeafSSL.git cd LeafSSL git init submodule
bash ./scripts/get_data.sh
bash ./scripts/get_weights.sh
python -m LeafSSL.train -c LeafSSL/config/sigreg_from_scratch.yaml
Edit the config file to point towards your local DinoV3 ViT-L checkpoint.
python -m LeafSSL.train -c LeafSSL/config/sigreg_dinov3_lora.yaml
python -m LeafSSL.test -c LeafSSL/config/test/plant_evaluation_full_val.yaml
python -m LeafSSL.test -c LeafSSL/config/test/plant_evaluation_full_test.yaml
If this project is useful to your academic work, please cite:
@inproceedings{becker2027wacv,
title = {{Self-Supervised Representation Learning for Few-shot Plant Variety Classification}},
author = {P.N. Becker and E.A. Marks and Y.-L. Chong and T. Kapler and J. Behley and L. Wiesmann and C. Stachniss},
year = 2027,
booktitle = wacv
}
If you have any questions, feel free to contact:
- Paul Nelson Becker {[email protected]}