dgyawa/yolo12n-seg-coco

YOLO12n Instance Segmentation — COCO Pretrained

This repository provides an unofficial community pretrained YOLO12n instance-segmentation checkpoint trained on the COCO dataset using the Ultralytics framework.

At the time of this release, Ultralytics provides the YOLO12 segmentation architecture but does not publish official pretrained YOLO12 segmentation .pt weights.

This checkpoint is an independent community release and is not an official Ultralytics model.

Model Details

Property	Value
Architecture	YOLO12n Segmentation
Task	Instance segmentation
Classes	80 COCO categories
Parameters	2,855,056
Training image size	640
Epochs	100
Batch size	16
Configured device	GPU 0
Stored parameter precision	FP16
Ultralytics version	8.4.18
Checkpoint date	March 4, 2026
Checkpoint	yolo12n-seg.pt
File size	6,109,860 bytes
Embedded license	AGPL-3.0

The training configuration records Ultralytics coco.yaml, and the checkpoint contains the standard 80 COCO category names.

Validation Results

The following validation metrics were recovered from the completed training checkpoint.

Metric	Bounding Boxes	Masks
Precision	67.508%	66.782%
Recall	49.684%	47.927%
mAP@50	55.282%	52.393%
mAP@50–95	39.513%	33.032%

Evaluation note: These are validation results stored in the completed training checkpoint. They have not yet been independently reproduced using a fresh post-release validation run.

Training Configuration

Setting	Value
Epochs	100
Image size	640
Batch size	16
Optimizer setting	auto
Seed	0
Deterministic training	Enabled
AMP	Enabled
Warmup	3 epochs
Cumulative logged training time	40.77 hours
Ultralytics	8.4.18

The specific optimizer selected internally by Ultralytics optimizer=auto was not retained in the available checkpoint metadata.

The training configuration records yolo12n.pt as the requested initialization. The available checkpoint metadata does not establish exactly which pretrained parameters were successfully transferred into the segmentation architecture. No stronger claim about initialization transfer is therefore made here.

Usage

Install the required packages:

pip install ultralytics huggingface_hub

Download and run the model:

from huggingface_hub import hf_hub_download
from ultralytics import YOLO

checkpoint = hf_hub_download(
    repo_id="dgyawa/yolo12n-seg-coco",
    filename="yolo12n-seg.pt",
)

model = YOLO(checkpoint)

results = model("image.jpg")

for result in results:
    print(result.boxes)
    print(result.masks)

Reproducibility Artifacts

This repository includes:

• checkpoint-metadata.json — metadata recovered from the checkpoint,
• training-history.csv — saved 100-epoch training history,
• classes.csv — all 80 stored COCO classes,
• SHA256SUMS — checksum for verifying the released checkpoint.

Intended Use

This checkpoint is intended primarily for transfer learning, downstream instance-segmentation initialization, YOLO12 segmentation research, and comparison of supervised and self-supervised pretraining approaches.

Users should independently validate the checkpoint for their target application before deployment.

Limitations

This is an unofficial community-trained checkpoint.

The exact training-image manifest is not encoded in the checkpoint and therefore cannot be independently reconstructed from the model artifact alone.

GPU model, standardized inference speed, FLOPs, and per-class AP were not retained in the available checkpoint metadata.

Acknowledgments

This checkpoint was trained using the Ultralytics YOLO software framework.

The YOLOv12 architecture originates from the work of Yunjie Tian, Qixiang Ye, and David Doermann.

The model was trained using the Microsoft COCO dataset introduced by Tsung-Yi Lin and colleagues.

This repository is an independent community release and is not affiliated with or endorsed by Ultralytics or the original YOLOv12 authors.

License

This checkpoint is distributed under the GNU Affero General Public License v3.0 (AGPL-3.0).

The checkpoint contains an embedded AGPL-3.0 license field and was produced using the Ultralytics training framework.

Users are responsible for ensuring that their use complies with applicable Ultralytics and AGPL-3.0 licensing requirements.

Citations

YOLOv12

@inproceedings{tian2025yolov12,
  title={YOLOv12: Attention-Centric Real-Time Object Detectors},
  author={Tian, Yunjie and Ye, Qixiang and Doermann, David},
  booktitle={Advances in Neural Information Processing Systems},
  year={2025}
}

Microsoft COCO

@inproceedings{lin2014microsoft,
  title={Microsoft COCO: Common Objects in Context},
  author={Lin, Tsung-Yi and Maire, Michael and Belongie, Serge J. and Hays, James and Perona, Pietro and Ramanan, Deva and Dollar, Piotr and Zitnick, C. Lawrence},
  booktitle={European Conference on Computer Vision},
  pages={740--755},
  year={2014},
  doi={10.1007/978-3-319-10602-1_48}
}

Attribution

If you use this checkpoint in research, please cite the original YOLOv12 work and Microsoft COCO paper and acknowledge the Ultralytics framework used for training.