| ---
|
| license: apache-2.0
|
| library_name: libreyolo
|
| tags:
|
| - object-detection
|
| - rt-detr
|
| - rt-detr-v4
|
| - detr
|
| ---
|
|
|
| # LibreRTDETRv4l
|
|
|
| RT-DETRv4-L HGNetv2 COCO detector, repackaged for the [LibreYOLO](https://github.com/LibreYOLO/libreyolo) framework.
|
|
|
| ## Source
|
|
|
| Derived from [https://github.com/RT-DETRs/RT-DETRv4](https://github.com/RT-DETRs/RT-DETRv4) at commit `43663fb103fc86a5da28806d8aeecdc707c4d6e0` and upstream checkpoint:
|
| `https://drive.google.com/file/d/1shO9EzZvXZyKedE2urLsN4dwEv8Jqa_8`.
|
|
|
| RT-DETRv4 is Apache-2.0 licensed. The upstream training recipe uses a DINOv3 teacher model; DINOv3 materials are not included in this repository.
|
|
|
| ## Modifications
|
|
|
| EMA weights were unwrapped, the training-only `encoder.feature_projector.*` tensors were omitted, and the checkpoint was saved with LibreYOLO metadata schema v1.0. The inference graph uses the exported student detector weights.
|
|
|
| ## Validation
|
|
|
| This file was validated locally with LibreYOLO checkpoint metadata schema v1.0, loaded through `LibreYOLO`, and evaluated on `coco128.yaml` with `workers=0`, `batch=8`, `conf=0.001`, `iou=0.6`.
|
|
|
| | Metric | Value |
|
| |---|---:|
|
| | COCO128 mAP50-95 | 0.6526 |
|
| | COCO128 mAP50 | 0.8398 |
|
| | SHA256 | `ffb706d4ec6f1d0d3ec11f3b22541bfcdb9b35982922ace4a558db563b2f3a12` |
|
| | File size | 120.2 MB |
|
|
|
| ## License
|
|
|
| Apache License 2.0. See the [`LICENSE`](./LICENSE) and [`NOTICE`](./NOTICE) files in this repository.
|
|
|
