FarmSAM student checkpoints

Compact TinyViT students for SAM 3's video tracker. A student replaces SAM 3's image backbone and learns from unlabelled frames of one farm, while the rest of the tracker stays the same. On dairy calves it keeps identity correct 99.23 % of the time (SAM 3: 99.80 %), with a backbone that is 91 % smaller and 3.6× faster.

You also need SAM 3's own weights from Meta, under the gate of facebook/sam3. Code and full results: github.com/Sonam525/FarmSAM. Preprint: arXiv 2604.27128.

Checkpoints

Each student: TinyViT-21M-512 trunk plus feature-pyramid head, 40.66 M backbone parameters, 60.20 M in total with the SAM 3 tracker. Input 1,024 × 1,024. Output map [1, 1024, 72, 72].

Student	File	Distilled on	Frames
Calf student (night-mined)	calf/calf_student_night_mined/best_backbone_student_v2.pt	Dairy calves, day and night	10,514
PigTrack student	pigtrack/pigtrack_student/best_backbone_student_pigtrack.pt	PigTrack	3,000
Edinburgh student	edinburgh_noncommercial/edinburgh_student/best_backbone_student.pt	Edinburgh	3,000

There is one student per dataset. Pick the one closest to your setting, and distil your own if your pens look different. Hashes are in CHECKSUMS.sha256.

How to load

git clone https://github.com/Sonam525/FarmSAM && cd FarmSAM
pip install -r requirements.txt
hf auth login      # with facebook/sam3 terms accepted

import sys, torch
from huggingface_hub import hf_hub_download
from PIL import Image

sys.path.insert(0, "benchmark/core")
import bench_core as core

path = hf_hub_download("Sonam5/FarmSAM", "calf/calf_student_night_mined/best_backbone_student_v2.pt")
student, meta = core.load_student(path, device="cuda")      # prefers the EMA weights, casts to bf16

x = core.make_student_transform()(Image.open("frame.jpg").convert("RGB"))
with torch.no_grad():
    features = student(x[None].to("cuda", torch.bfloat16))  # [1, 1024, 72, 72]

core.load_arm swaps the student into Sam3TrackerVideoModel. The tracker's neck, mask decoder, memory attention and memory encoder stay unchanged.

.pt is pickle, so check the SHA-256 before you load a file. Use the EMA weights; load_student picks them for you.

Results

Each row uses a student distilled on that dataset. Tracks start from reference boxes or masks at the first frame.

Dataset, metric	Student	SAM 3	SAM 2.1-small	SAM 2.1-tiny	EdgeTAM
Dairy calves, identity-correct %, 6 windows	99.23	99.80	98.77	98.96	91.29
Edinburgh pigs, HOTA, 4 clips	74.05	75.54	75.18	74.52	71.21
PigTrack, HOTA, 25 sequences	73.25	85.87	80.90	79.38	71.16

Model	Backbone parameters	Total parameters	Backbone, ms per frame	End to end, fps
Student	40.66 M	60.20 M	41.56–42.61	4.83–5.01
SAM 3	446.24 M	465.78 M	152.46	3.25
SAM 2.1-small	33.95 M	46.06 M	15.04	7.83
SAM 2.1-tiny	26.85 M	38.96 M	12.78	8.00
EdgeTAM	4.71 M	13.90 M	10.03	13.05

One A10 GPU, 4 objects, frame decoding included. Calf identity is scored against expert-verified identity chains on masks from an earlier run of the calf student, so it measures how well a track stays on its animal rather than mask quality. The speed gain shrinks as more objects are tracked. Full tables: RESULTS.md.

Intended use

Research on feature distillation and on tracking animals from fixed farm cameras. Evaluate a student on your own pen before you rely on it. Tracks need initial boxes, masks or a detector.

Not intended for decisions about individual animals without review, for tracking people, or for any use the SAM License excludes. The Edinburgh student is for non-commercial use only.

Terms

The code on GitHub is MIT, except parts under other terms. These weights are not. All students are released under the SAM License, and the two Edinburgh students carry a non-commercial limit on top.

• The students were trained to match SAM 3 backbone features and run inside SAM 3's tracker, so we apply the SAM License to be safe. If you pass a student on, you must do so under those terms and include a copy of LICENSE_SAM.txt. Its trade control limits apply.
• Non-commercial limit. edinburgh_noncommercial/ holds a student that may be used for non-commercial purposes only. It was distilled on frames of the Edinburgh pig dataset (CC BY-NC; Bergamini et al. 2021). This limit comes from the data licence and applies on top of the SAM License. Please cite that dataset.
• The PigTrack student was distilled on PigTrack (CC BY 4.0). Please credit it.
• The calf student was distilled on Calf-Play-Behavior-Dataset, FarmA_R1 (CC0 1.0), days 18 and 19 April only.
• The trunk starts from timm/tiny_vit_21m_512.dist_in22k_ft_in1k, changed through training.

See NOTICE.md. This note is not legal advice.

Citation

@misc{yang2026distillation,
  title         = {Lightweight Distillation of {SAM}~3 and {DINOv3} for Edge-Deployable
                   Individual-Level Livestock Monitoring and Longitudinal Visual Analytics},
  author        = {Yang, Haiyu and Hostens, Miel},
  year          = {2026},
  eprint        = {2604.27128},
  archivePrefix = {arXiv},
  url           = {https://arxiv.org/abs/2604.27128}
}

If you publish results from a student, you must acknowledge SAM 3: clause 1.b.ii of the SAM License requires it.