yolo26s-pose-onnx

YOLO26s pose (ONNX) for qtpose

ONNX export of the Ultralytics YOLO26s-pose model (COCO-Pose, 17 keypoints), published for use by qtpose. qtpose uses it as a single-stage person + pose model and, as yolo26s_person, as the person detector for two-stage pose models.

Licence

Ultralytics YOLO26 weights are released under the GNU AGPL-3.0 licence, and this derived ONNX file is distributed under the same licence. See https://www.ultralytics.com/license for the Ultralytics licence terms, including the enterprise licence option.

Source checkpoint

Ultralytics yolo26s-pose.pt (COCO-Pose), from the Ultralytics assets releases.

Export

Exported with Ultralytics 8.4.15 (PyTorch 2.10.0) on 2026-02-24, as recorded in the file's embedded metadata, equivalent to:

from ultralytics import YOLO

YOLO("yolo26s-pose.pt").export(
    format="onnx", imgsz=640, opset=12, simplify=True, dynamic=False, nms=False
)

The YOLO26 head is end-to-end (end2end=True), so no post-processing NMS is needed. qtpose's helper scripts/dev/convert_yolo26_to_onnx.py --task pose --variant s --validate reproduces this kind of export.

Model file

• yolo26s_pose.onnx, 41858651 bytes
• Input: images, float32, shape (1, 3, 640, 640) (RGB, 0-1 scaled, letterboxed)
• Output: output0, float32, shape (1, 300, 57). Each row is [x1, y1, x2, y2, score, class_id, kx1, ky1, kc1, …, kx17, ky17, kc17] in input-image pixels; keypoints follow the COCO-17 order (nose, eyes, ears, shoulders, elbows, wrists, hips, knees, ankles; left before right).
• Classes: person (0).
• SHA-256: af718721825e4795ad5db2740a776455fa991064ff6cf837f95e825a92fa8e0b