Control-R1: Doc-layout object detection (fine-tuned DETR)

This model is a fine-tuned DETR (ResNet-50) object-detection model for document / newspaper page-layout analysis. It detects 7 region types in scanned historical newspaper pages and was trained on the biglam/loc_beyond_words dataset.

Base model & license

• Base model: facebook/detr-resnet-50
• License: Apache-2.0 (open license — free to share, modify and use commercially)
• The fine-tuned weights in this repository inherit the Apache-2.0 license.

Classes (7)

Photograph, Illustration, Map, Comics/Cartoon, Editorial Cartoon, Headline, Advertisement

Training data

• Dataset: biglam/loc_beyond_words
• Train: 2846 images
• Validation: 712 images
• COCO-style annotations (xywh bounding boxes).

Training details

• Model: facebook/detr-resnet-50, all parameters fine-tuned (backbone unfrozen)
• Epochs: 4
• Batch size: 4 (gradient accumulation = 2)
• Optimizer: AdamW (lr = 0.0001, weight decay on non-bias/LN params)
• LR schedule: linear warmup + cosine decay
• Image preprocessing: resize to shortest edge = 800px, longest = 1333px (aspect preserved), ImageNet normalization
• Mixed precision: bf16/fp16 autocast (GPU)
• Hardware: Hugging Face Jobs GPU

Validation results (COCO eval on the 712-image validation split)

Metric	Value
mAP (IoU 0.50:0.95)	0.1658
AP @ IoU 0.50	0.2761
AP @ IoU 0.75	0.1802
AR (max 100 dets)	0.2908

Per-class AP (IoU 0.50:0.95):

• Photograph: 0.3484
• Illustration: 0.0000
• Map: 0.0000
• Comics/Cartoon: 0.0385
• Editorial Cartoon: 0.0000
• Headline: 0.4418
• Advertisement: 0.3321

Quick usage

from transformers import AutoModelForObjectDetection, AutoImageProcessor
from PIL import Image

model = AutoModelForObjectDetection.from_pretrained("harness-race/control-r1")
processor = AutoImageProcessor.from_pretrained("harness-race/control-r1")

image = Image.open("page.png").convert("RGB")
inputs = processor(images=image, return_tensors="pt")
outputs = model(**inputs)
results = processor.post_process_object_detection(outputs, target_sizes=[(image.height, image.width)], threshold=0.4)
for r in results[0]:
    print(model.config.id2label[int(r['labels'])], round(r['scores'].item(),3) if hasattr(r['scores'],'item') else r['scores'], [round(c,1) for c in r['boxes'].tolist()])

Intended use & limitations

Trained for research on historical newspaper layout analysis. Best on page layouts similar to the loc_beyond_words training distribution; large format/styled pages not seen in training may be missed. Evaluation was done on the dataset's official 712-image validation split; run at 2026-08-07 (UTC) on HF Jobs.

---

Control-R1 — a layout-model entry. Trained on Hugging Face Jobs (< $5 budget).