adopd/YOLOv12-X-detection-ADOPD

YOLOv12-X Detection ADOPD

Thinking with Anchors Project | ADOPD 2026 Paper: Thinking with Anchors: Grounded and Efficient Document Reasoning | ADOPD 2024 Paper | Dataset | Code

Model Overview

• Model developer: Thinking with Anchors project contributors
• Base architecture: YOLOv12-X
• Model type: one-stage object detector
• Task: grouped document text-region detection (Doc2Box)
• Fine-tuning dataset: adopd/adopd2026
• Input: RGB document image
• Output: text-region boxes, confidence scores, and class IDs

Description

YOLOv12-X Detection ADOPD is a YOLOv12-X model fine-tuned for single-class grouped text-region detection in document images.

Training Data

Public Doc2Box supervision is stored in:

ocr.grouped_blocks[].bbox_xyxy

The companion exporter writes one normalized YOLO box per grouped OCR region with the class name text.

Checkpoint Format

model.pt is a native Ultralytics checkpoint. It requires the pinned YOLOv12 source containing the A2C2f module; generic Ultralytics installations may not deserialize it.

Quick Start

git clone https://github.com/SichenZhu/ADOPD2026.git
cd ADOPD2026/release_code

git clone https://github.com/sunsmarterjie/yolov12.git upstream/yolov12
git -C upstream/yolov12 checkout 01a22c0603e0eaa6d9bd62120a391e744d92cea2

python -m pip install -e model_zoo/common
python -m pip install -e upstream/yolov12
python -m pip install -e model_zoo/yolov12

hf download adopd/YOLOv12-X-detection-ADOPD \
  --local-dir checkpoints/yolov12-text

adopd-yolo-infer \
  --checkpoint checkpoints/yolov12-text/model.pt \
  --image document.jpg \
  --confidence 0.25 \
  --output prediction.json

The output JSON contains pixel-space boxes, confidence scores, class IDs, and image size.

Fine-Tuning And Evaluation

Prepare train and validation with adopd-yolo-prepare --task detect, then train with adopd-yolo-train --task detect --data PREPARED_ROOT/detect.yaml. See yolov12 for complete commands.

Limitations

This model detects grouped text regions but does not transcribe text. It is single-class and confidence thresholds should be calibrated on the deployment domain.

License

This checkpoint is released under AGPL-3.0. See the included license.

Citation

Please cite the ADOPD 2026 and ADOPD 2024 papers.

@misc{zhu2026thinkingwithanchors,
  title={Thinking with Anchors: Grounded and Efficient Document Reasoning},
  author={Sichen Zhu and Yuchen Zhu and Wenzhuo Xu and Jason Kuen and Wanrong Zhu and Jing Shi and Xuan Shen and Quanyi Wang and Yiwei Wang and Yujun Cai and Bing Shuai and Qin Zhang and Yongxin Chen and Shilong Liu and Molei Tao and Jiuxiang Gu},
  year={2026}
}

@inproceedings{gu2024adopd,
  title={{ADOPD}: A Large-Scale Document Page Decomposition Dataset},
  author={Jiuxiang Gu and Xiangxi Shi and Jason Kuen and Lu Qi and Ruiyi Zhang and Anqi Liu and Ani Nenkova and Tong Sun},
  booktitle={The Twelfth International Conference on Learning Representations},
  year={2024},
  url={https://openreview.net/forum?id=x1ptaXpOYa}
}