---
license: mit
library_name: libreyolo
tags:
  - object-detection
  - yolov9
  - nms-free
  - end-to-end
---

# LibreYOLO9E2Ec

YOLOv9 end-to-end (NMS-free) variant, size-c, for LibreYOLO.

## Source

Backbone, neck and one-to-many head derived from
[MultimediaTechLab/YOLO](https://github.com/MultimediaTechLab/YOLO)
(release v1.0-alpha, file `v9-c.pt`).
Copyright (c) 2024 Kin-Yiu Wong and Hao-Tang Tsui. Licensed under the MIT License.

## Modifications

Backbone, neck and one-to-many head: state-dict key remapping only;
learned parameters unchanged from upstream.

One-to-one head (`head.one2one_cv2`, `head.one2one_cv3`): trained
from scratch on COCO 2017 with the rest of the network frozen.

mAP (COCO val2017, COCOeval): **mAP50-95 = 0.4944, mAP50 = 0.6603**.

See [LibreYOLO issue #84](https://github.com/LibreYOLO/libreyolo/issues/84)
and [PR #156](https://github.com/LibreYOLO/libreyolo/pull/156) for the recipe.

## License

MIT License. See the [`LICENSE`](./LICENSE) file in this repository.

## Acknowledgements

The YOLOv9-E2E architecture (NMS-free one-to-one head added to YOLOv9) and the
frozen-backbone fine-tuning recipe used to train the released weights are the
work of [testdummyvt](https://github.com/testdummyvt) . first proposed in
[LibreYOLO issue #84](https://github.com/LibreYOLO/libreyolo/issues/84) and
contributed via [PR #156](https://github.com/LibreYOLO/libreyolo/pull/156).