Instructions to use sdoerrich97/true_colon_rtdetr_realcolon_s0 with libraries, inference providers, notebooks, and local apps. Follow these links to get started.
- Libraries
- ultralytics
How to use sdoerrich97/true_colon_rtdetr_realcolon_s0 with ultralytics:
from huggingface_hub import hf_hub_download from ultralytics import YOLO # pick the weights file from this repo's "Files and versions" tab weights = hf_hub_download("sdoerrich97/true_colon_rtdetr_realcolon_s0", "<weights>.pt") model = YOLO(weights) source = 'http://images.cocodataset.org/val2017/000000039769.jpg' model.predict(source=source, save=True) - Notebooks
- Google Colab
- Kaggle
Model card: canonical template, code authors, arXiv citation
Browse files
README.md
CHANGED
|
@@ -12,27 +12,35 @@ pipeline_tag: object-detection
|
|
| 12 |
|
| 13 |
# TRUE-Colon — RT-DETR (REAL-Colon, seed 0)
|
| 14 |
|
| 15 |
-
|
| 16 |
-
|
| 17 |
-
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
|
| 18 |
|
| 19 |
This is one of three seeds. Reporting a single seed of this model in isolation would
|
| 20 |
overstate its precision; the paper aggregates all three.
|
| 21 |
|
| 22 |
## Model Details
|
| 23 |
|
| 24 |
-
- **Developed by:**
|
|
|
|
| 25 |
- **Model type:** Single-class object detector (`lesion`)
|
| 26 |
- **License:** AGPL-3.0, inherited from Ultralytics. This is **not** the MIT license of the
|
| 27 |
-
`true_colon` evaluation package
|
| 28 |
-
source-availability obligations.
|
| 29 |
- **Framework:** Ultralytics
|
| 30 |
|
| 31 |
### Model Sources
|
| 32 |
|
| 33 |
- **Paper:** https://arxiv.org/abs/2608.13711
|
| 34 |
- **Repository:** https://github.com/sdoerrich97/true-colon
|
| 35 |
-
- **
|
| 36 |
|
| 37 |
### Architecture and training
|
| 38 |
|
|
@@ -118,19 +126,19 @@ card, because a single seed's values are not the quantity the paper reports.
|
|
| 118 |
## Citation
|
| 119 |
|
| 120 |
```bibtex
|
| 121 |
-
@
|
| 122 |
-
title
|
| 123 |
-
author
|
| 124 |
-
|
| 125 |
-
|
| 126 |
-
|
| 127 |
-
|
| 128 |
-
|
| 129 |
-
primaryClass = {eess.IV},
|
| 130 |
-
doi = {10.48550/arXiv.2608.13711}
|
| 131 |
}
|
| 132 |
```
|
| 133 |
|
| 134 |
-
## Contact
|
| 135 |
|
| 136 |
-
|
|
|
|
|
|
|
|
|
| 12 |
|
| 13 |
# TRUE-Colon — RT-DETR (REAL-Colon, seed 0)
|
| 14 |
|
| 15 |
+
Official weights for **TRUE-Colon: Exposing a Consistent Transfer Asymmetry in Real-Time
|
| 16 |
+
Polyp Detection** (MICCAI 2026 EndoLINA Workshop). This repo holds the **`rtdetr`**
|
| 17 |
+
detector at **seed 0**, trained on REAL-Colon: 60 complete, unedited colonoscopy
|
| 18 |
+
procedures rather than curated lesion-centric clips.
|
| 19 |
+
|
| 20 |
+
TRUE-Colon is an evaluation protocol first. Its finding is that detectors trained on
|
| 21 |
+
curated benchmarks collapse on full procedures, while procedure-trained models like this
|
| 22 |
+
one suppress non-polyp content and still transfer back to curated data. The detector itself
|
| 23 |
+
is an off-the-shelf RT-DETR: what is released here is a baseline trained under the
|
| 24 |
+
paper's protocol, not a new architecture.
|
| 25 |
|
| 26 |
This is one of three seeds. Reporting a single seed of this model in isolation would
|
| 27 |
overstate its precision; the paper aggregates all three.
|
| 28 |
|
| 29 |
## Model Details
|
| 30 |
|
| 31 |
+
- **Developed by:** Sebastian Doerrich, Andreas Franz Schwab (xAILab, University of Bamberg)
|
| 32 |
+
- **Funded by:** Hightech Agenda Bayern (HTA) of the Free State of Bavaria, Germany
|
| 33 |
- **Model type:** Single-class object detector (`lesion`)
|
| 34 |
- **License:** AGPL-3.0, inherited from Ultralytics. This is **not** the MIT license of the
|
| 35 |
+
`true_colon` evaluation package (© Sebastian Doerrich, Andreas Franz Schwab). Using these
|
| 36 |
+
weights in a product carries the AGPL's source-availability obligations.
|
| 37 |
- **Framework:** Ultralytics
|
| 38 |
|
| 39 |
### Model Sources
|
| 40 |
|
| 41 |
- **Paper:** https://arxiv.org/abs/2608.13711
|
| 42 |
- **Repository:** https://github.com/sdoerrich97/true-colon
|
| 43 |
+
- **Package:** https://pypi.org/project/true-colon/
|
| 44 |
|
| 45 |
### Architecture and training
|
| 46 |
|
|
|
|
| 126 |
## Citation
|
| 127 |
|
| 128 |
```bibtex
|
| 129 |
+
@article{doerrich2026truecolon,
|
| 130 |
+
title={TRUE-Colon: Exposing a Consistent Transfer Asymmetry in Real-Time Polyp Detection},
|
| 131 |
+
author={Sebastian Doerrich and Andreas Franz Schwab and Francesco {Di Salvo} and Shyam Nandan Rai and Hanh Huyen My Nguyen and Christian Ledig},
|
| 132 |
+
year={2026},
|
| 133 |
+
eprint={2608.13711},
|
| 134 |
+
archivePrefix={arXiv},
|
| 135 |
+
primaryClass={eess.IV},
|
| 136 |
+
url={https://arxiv.org/abs/2608.13711},
|
|
|
|
|
|
|
| 137 |
}
|
| 138 |
```
|
| 139 |
|
| 140 |
+
## Model Card Contact
|
| 141 |
|
| 142 |
+
For questions or issues, please open an issue in the
|
| 143 |
+
[GitHub repository](https://github.com/sdoerrich97/true-colon), or contact
|
| 144 |
+
sebastian.doerrich@uni-bamberg.de.
|