LibreRFDETRn-obb
Ultra-experimental LibreYOLO RF-DETR-nano oriented object detection weights fine-tuned for vehicle OBB detection.
These weights are ultra-experimental development weights. They were produced while validating LibreYOLO OBB training support and should not be treated as production or benchmark-official weights.
Source
Initialized from LibreRFDETRn detect weights, derived from RF-DETR / LW-DETR lineage and distributed by LibreYOLO with Apache-2.0-compatible metadata.
Fine-tuned on
My First Project - v8, a Roboflow Universe UAV vehicle OBB dataset provided by a Roboflow user:
https://universe.roboflow.com/bobo-48pem/my-first-project-ewwrmThis model was not trained on DOTA.
Dataset
Dataset license: Creative Commons Attribution 4.0 International (CC BY 4.0).
Classes:
- bike
- bus
- car
- other_vehicle
- taxi
- truck
Local development export metadata reported 932 source/export images, YOLOv8 Oriented Object Detection format, and Roboflow preprocessing/augmentation.
Metrics
UAV-OBB validation split, best epoch 10:
- mAP50: 0.719320
- mAP50-95: 0.401842
- mAP75: 0.406394
- precision: 0.066207
- recall: 0.836576
Additional independent checks from development:
- PyTorch test split: mAP50=0.7350, mAP50-95=0.4453
- ONNX backend test split: mAP50=0.7356, mAP50-95=0.4451
Modifications
The checkpoint is a lean LibreYOLO inference checkpoint produced from the best EMA training checkpoint. Optimizer, resume config, raw train model state, and EMA resume buffers were removed before upload.
License
These fine-tuned weights are released under CC BY 4.0 to preserve the attribution requirements of the training dataset. See
LICENSE and
NOTICE.
