LibreYOLO9s-pose

This repository contains LibreYOLO9s-pose.pt, an EXTREMELY experimental LibreYOLO9-s pose-estimation checkpoint for COCO-17 human keypoints.

Experimental Status

These weights are published as an early preview so LibreYOLO users can test the new task="pose" path. They are not a final production model, may be replaced, and may change without compatibility guarantees.

Validation Snapshot

Training data: COCO person keypoints
Task: single-class person pose, 17 keypoints, kpt_shape: [17, 3]
Run: 30 epochs, 640 image size
Best checkpoint: epoch 30
COCO val keypoints AP50-95: 0.5739
COCO val keypoints AP50: 0.8310

Usage

from libreyolo import LibreYOLO

model = LibreYOLO("yolo9-s-pose")
result = model("image.jpg")
print(result.keypoints)

Autodownload in LibreYOLO emits an experimental warning for this checkpoint.

Source

Trained by LibreYOLO contributors with the LibreYOLO pose-training implementation, initialized from the LibreYOLO9-s detection checkpoint and fine-tuned on COCO person keypoints.

Modifications

This is a native LibreYOLO training checkpoint with LibreYOLO checkpoint metadata. It is not a direct upstream conversion.

License

MIT License. See the LICENSE file in this repository.

Downloads last month: -; Downloads are not tracked for this model. How to track