SYNTOM-SAM3: SAM 3 fine-tuned for tomato organ segmentation

SAM 3 fully fine-tuned on SYNTOM, for text-conditioned segmentation of four tomato organs: leaf, stem, flower, fruit.

This is the Full FT SAM 3 model from Text-conditioned Segmentation for Tomato Phenotyping via Procedural Synthetic Data, the variant that scores best on real greenhouse imagery in the paper.

Trained only on synthetic data, evaluated on real photographs.

Results

Fruit IoU on three real-world tomato datasets, predictions reduced to fruit vs non-fruit:

model	TomatoMAP	Laboro	Rob2Pheno	macro	image-weighted
this model	0.7785	0.7371	0.4529	0.6562	0.7372
zero-shot SAM 3	0.5938	0.6875	0.2995	0.5269	0.6210

Fine-tuning on synthetic data lifts macro fruit IoU by 13 points over the zero-shot baseline, without a single real training image.

Files

sam3_fullft_syntom.pt      the weights, 3.1 GiB
provenance.json            file size and checksum
LICENSE                    the SAM License, which governs these weights

Loading

import torch

sd = torch.load("sam3_fullft_syntom.pt", map_location="cpu", weights_only=True)["model"]
model.detector.load_state_dict(sd, strict=False)

The keys address SAM 3's detector and use the research code's parameter names, so load them with the sam3 package rather than transformers. Meta publishes SAM 3 in both forms. This fine-tune follows the sam3.pt one.

License

These weights are a derivative work of Meta's SAM 3 and are distributed under the SAM License, a copy of which is in LICENSE. They are not covered by the CC BY 4.0 licence on the SYNTOM dataset, which is a separate artifact.

They are provided "as is", without warranties or conditions of any kind, express or implied, and without any guarantee of fitness for a particular purpose. You are solely responsible for determining whether they are appropriate for your use and assume all risk arising from it.

Section 1(b)(ii) of the SAM License requires that published research using SAM Materials acknowledge them, so cite SAM 3 alongside this work.

Citation

Accepted at the 11th Workshop on Computer Vision in Plant Phenotyping and Agriculture (CVPPA), in conjunction with the European Conference on Computer Vision (ECCV), Malmo, Sweden, September 2026.

@inproceedings{mounir2026syntom,
  title         = {Text-conditioned Segmentation for Tomato Phenotyping via
                   Procedural Synthetic Data},
  author        = {Mounir Samy, Cieslak Mikolaj, Dhieb Najmeddine,
                   Ghazzai Hakim, Klein Jonathan, Froehlich Katja,
                   Pirk Soeren, Palubicki Wojciech, Setti Gianluca,
                   Eltawil Ahmed M., Michels Dominik L.},
  booktitle     = {11th Workshop on Computer Vision in Plant Phenotyping and
                   Agriculture (CVPPA), in conjunction with the European
                   Conference on Computer Vision (ECCV)},
  address       = {Malmo, Sweden},
  month         = sep,
  year          = {2026},
  eprint        = {2607.18576},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CV}
}

Acknowledgment

This work was supported by funding from King Abdullah University of Science and Technology (KAUST) - Center of Excellence on Sustainable Food Security, under award number 5934.

Contact

• Samy Mounir, samy.mounir@kaust.edu.sa
• Mikolaj Cieslak, mikolaj.cieslak@kaust.edu.sa
• Najmeddine Dhieb, najmeddine.dhieb@kaust.edu.sa