Samuka007/sam3.1-plant-fewshot

SAM3.1 Plant Few-Shot

A single-image few-shot fine-tuned SAM3.1 image model for plant instance segmentation.

Intended Use

This checkpoint is an experimental pre-labeling model for the plant dataset used in the accompanying training run. It is suitable for visual inspection and iterative annotation assistance, not as a claim of reliable cross-scene generalization.

The model was trained on one 2000x2000 image containing 91 plant polygon instances. The seven CVAT point annotations in the test archive were intentionally excluded from training and evaluation.

Usage

This is an official SAM3 image trainer checkpoint containing the full image-model state under checkpoint.pt. It is not a Transformers checkpoint. Use the official facebookresearch/sam3 code at commit 96914d2425f90a64f45ca977c2b5165418099543, load the gated base checkpoint facebook/sam3.1/sam3.1_multiplex.pt, then apply the model state from this repository's checkpoint.pt with strict loading.

The exact fine-tuning model factory, Hydra configs, and inference renderer used for this run are included under pipeline/. Adjust the absolute dataset and checkpoint paths in the Hydra config for your environment. The upstream base checkpoint is intentionally not mirrored in this repository.

Training

• Base model: facebook/sam3.1
• Base revision: daa63191845a41281374e725f4c9e51c7a824460
• Upstream code commit: 96914d2425f90a64f45ca977c2b5165418099543
• Training steps: 400
• Resolution: 1008
• Batch size: 1
• Precision: bfloat16
• Prompt: plant
• Frozen: language backbone and pretrained vision tower; the missing SAM3.1 fourth FPN projection and detection/segmentation heads were trainable

License

The upstream SAM3.1 checkpoint and derivative weights are distributed under the Meta SAM License. The license text is included as LICENSE. By downloading or using this repository, you are responsible for complying with that license and any applicable upstream access requirements.

Versioning

The main branch points to the current recommended iteration. Each promoted iteration is also published as an immutable tag, beginning with v0.1.0. Do not rely on main for reproducibility; pin a tag or commit.