control-r3 — DETR object detection for news-page layout regions

This model is a fine-tuned DetrForObjectDetection (DETR with a ResNet-50 backbone, starting from facebook/detr-resnet-50) tuned to localize layout regions in historical and modern newspaper pages. It predicts bounding boxes for the 7 classes of the BigLam “Locating Objects Beyond Words” dataset:

class id	label
0	Photograph
1	Illustration
2	Map
3	Comics/Cartoon
4	Editorial Cartoon
5	Headline
6	Advertisement

The base model facebook/detr-resnet-50 is released under the Apache 2.0 license, so this fine-tune can be freely shared and used.

Model Details

• Model type: DETR (DEtection TRansformer), encoder-decoder Transformer with a ResNet-50 backbone.
• Backbone: ResNet-50 (provided by timm), pretrained on ImageNet.
• Number of queries: 100.
• Input: RGB image resized/padded to a square of 600 × 600 after a smallest-max-size resize.
• Output: per-image box predictions with class logits for the 7 layout classes (plus the implicit no object class).
• Optimized by: fine-tuned on the training split (2,846 images) of biglam/loc_beyond_words.
• Base model license: Apache 2.0.

Intended Use

The model is intended for document / newspaper-page layout analysis: given a scan or a page image, it detects coarse layout regions such as headlines, photographs/illustrations, advertisements, maps and comics. It is designed as a layout-region detector and is not meant for fine-grained text recognition or OCR (use OCR/HTR systems for reading text).

Example inference:

from transformers import pipeline

detector = pipeline("object-detection", model="harness-race/control-r3")
results = detector("path/to/newspaper_page.png")
# results: list of {label, score, box: {xmin, ymin, xmax, ymax}}

Training Data

• Dataset: biglam/loc_beyond_words (BigLam “Locating Objects Beyond Words”, a Library-of-Congress-derived newspaper layout dataset).
• Training split: 2,846 images.
• Validation split: 712 images (reported results are on this held-out split).
• Format: COCO-style bounding boxes [x, y, width, height] in pixel coordinates.
• The data was repackaged (only renaming annotation fields for compatibility) into harness-race/loc_beyond_words_coco; no annotations were modified.

Training Procedure

The model was fine-tuned end-to-end (all weights trainable) with the Hugging Face Trainer-style loop on a single NVIDIA T4 GPU (fp16 AMP), with the RGB images resized and padded to 600 × 600 and light augmentation (horizontal flip, random brightness/contrast, hue/saturation, random crop with box clipping).

Hyperparameters

• optimizer: AdamW (betas 0.9/0.999, eps 1e-8)
• learning rate: 1e-4 with linear schedule + 10% warmup
• weight decay: 1e-4
• train batch size: 4
• eval batch size: 8
• epochs: 8
• mixed precision: fp16 (native AMP)
• image size: 600 × 600 (resize + pad)
• gradient clipping: 1.0
• seed: 42

Evaluation is run at the end of every epoch and the checkpoint with the best validation mAP is kept.

Evaluation Results

Reported on the validation split (712 images), using COCO-style metrics (torchmetrics.MeanAveragePrecision, box_format=xyxy). Metrics are in %:

Metric	Value
mAP (IoU .5:.95)	24.65
mAP @ IoU 0.50	34.76
mAP @ IoU 0.75	28.39
mAR@100	37.94

Per-class mAP (IoU .5:.95):

Class	mAP
Photograph	39.02
Illustration	1.17
Map	0.03
Comics/Cartoon	13.21
Editorial Cartoon	0.00
Headline	59.51
Advertisement	59.64

The model detects Headline and Advertisement regions very well (>59 mAP) and detects Photograph and Comics/Cartoon reasonably. The rare classes (Illustration, Map, Editorial Cartoon) show very low mAP, which is largely a consequence of heavy class imbalance in the dataset (e.g. only ~215 Map and ~293 Editorial Cartoon object instances across the whole dataset vs ~27.9k Headline instances). More data or class-balancing/oversampling for those classes would improve them.

The raw per-epoch metrics are stored in val_metrics.json in this repository.

Bias, Risks, and Limitations

• Performance is strongly skewed by class imbalance; rare layout classes are unreliable.
• Detector quality on page layouts/engravings/styles outside the training distribution may degrade.
• Predictions should be reviewed when used in archival/curation workflows (layout regions may span multiple semantic zones; bounding boxes are approximate).
• This is a layout-region detector, not an OCR/HTR model.

Technical Notes

• Architecture: DetrForObjectDetection, 100 queries, 6 encoder + 6 decoder layers, d_model=256.
• Repository files: config.json, model.safetensors, preprocessor_config.json (DetrImageProcessorFast, size=600), val_metrics.json, train_detr.py (training script).

Acknowledgements

Based on the DETR model (Carion et al., 2020) and the Transformers library. Dataset from BigLam and the Library of Congress newspaper collections.