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

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

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 = 0.4198, mAP50 = 0.5741.

See LibreYOLO issue #84 and PR #156 for the recipe.

MIT License. See the LICENSE file in this repository.

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 . first proposed in LibreYOLO issue #84 and contributed via PR #156.