LibreRFDETRl-pose

LibreRFDETRl-pose

EXTREMELY experimental RF-DETR-l pose checkpoint for LibreYOLO.

This is a COCO-17 human pose preview checkpoint for LibreYOLO's task="pose" RF-DETR path. It is useful for testing and bootstrapping, but it is not a final benchmark release.

Checkpoint

• File: LibreRFDETRl-pose.pt
• Family: LibreRFDETR
• Size: l
• Task: pose
• Classes: person only
• Keypoints: COCO-17, (x, y, visibility)
• Validation image size: 704
• Additional training epochs for this checkpoint: 0

Initialization Method

Native RF-DETR-l detection checkpoint plus shared tensors from the trained LibreRFDETRs-pose checkpoint. The extra final decoder layer was initialized from the trained small-pose final decoder layer.

This method keeps the size-specific detection backbone and resolution-dependent tensors, then transfers the pose-specialized shared tensors from the small pose checkpoint. The checkpoint should still be treated as experimental until a full per-size training run is published.

COCO Keypoint Validation

Validation was run on COCO person keypoints val2017 through LibreYOLO's pose validator.

The validation artifacts are included as validation_metrics.json. Initialization details are included as initialization_summary.json.

Usage

from libreyolo import LibreRFDETR

model = LibreRFDETR("LibreRFDETRl-pose.pt", task="pose")
results = model.predict("image.jpg", imgsz=704)
print(results[0].keypoints)

Autodownload in LibreYOLO emits an experimental warning for this checkpoint.

Caveats

• Experimental checkpoint, not a final benchmark release.
• No additional fine-tuning epochs were run for this per-size checkpoint after transfer initialization.
• Pose export/runtime backends may have separate support status from PyTorch inference.
• Metrics are from LibreYOLO PR development artifacts, not from an independent external benchmark suite.