figpanel-yolov12

YOLOv12 model for detecting sub-panels and caption labels in scientific figures.

Built by Plottie — the scientific plot discovery platform.

Model Description

This model detects two classes of objects in composite scientific figures:

Class	Description
subplot	Individual plot/panel region within a composite figure
caption	Single-character label (a, b, c...) identifying each panel

The model was trained on 5,000+ annotated figures from open-access scientific journals spanning biology, medicine, physics, and engineering.

Property	Value
Architecture	YOLOv12
Input	Any image (PNG, JPEG, etc.)
Model Size	~18 MB
Classes	2 (subplot, caption)
Framework	Ultralytics

Examples

Blue boxes = detected subplots, Green boxes = detected captions. Images from open-access papers (CC BY 4.0).

How to Use

Install the Python package:

pip install figpanel

import figpanel

# Detect subplots and captions
results = figpanel.detect("figure.png")
# {'subplots': [(x1,y1,x2,y2,conf), ...], 'captions': [...]}

# Visualize detections
figpanel.visualize("figure.png", save="annotated.png")

For the full pipeline (detect + OCR + match + crop):

pip install figpanel[full]

panels = figpanel.extract("figure.png", "output/")
for panel in panels:
    print(f"Panel {panel.label}: confidence={panel.confidence:.2f}")

The model weights are downloaded automatically on first use via huggingface_hub.

Training Data

The model was trained on a custom dataset of 5,000+ annotated composite figures collected from open-access scientific publications (CC BY 4.0). Annotations include bounding boxes for subplot regions and single-character caption labels. The dataset covers diverse figure layouts across multiple scientific disciplines.

Limitations

• Optimized for composite figures with clearly separated panels; may underperform on continuous plots or single-panel figures
• Caption OCR expects single-character labels (a, b, c...); multi-character or numeric labels are not currently supported
• Performance may vary on figures with non-standard layouts (e.g., overlapping panels, inset plots)
• Trained primarily on figures from biology and methods journals; other domains may have lower accuracy

About Plottie

Plottie is the scientific plot discovery platform. We help researchers explore, collect, and find inspiration from high-quality scientific plots across open-access literature.

Citation

@software{figpanel,
  title = {figpanel: Scientific Figure Panel Detector},
  author = {Plottie},
  url = {https://github.com/Plottie/figpanel},
  version = {0.1.0},
  year = {2026},
  note = {Built by Plottie (https://plottie.art)}
}

License

AGPL-3.0 — inherited from Ultralytics. Academic and research use is unaffected.