README.md

metadata

license: apache-2.0
pretty_name: DigitizerBench
tags:
  - benchmark
  - plot-digitization
  - scientific-figures
  - synthetic
  - computer-vision
size_categories:
  - 1K<n<10K

DigitizerBench

A benchmark for recovering numerical curves from line-plot images

Project website · GitHub · Paper PDF · Online appendix

Each PNG comes with the original curve coordinates and the rendered tick positions. Use them to evaluate curve recovery and axis calibration separately.

Download and quick start

Choose Lite for a 60-image human-guided comparison or a quick pipeline check. Choose Full for automatic evaluation and analysis across rendering and signal conditions. Both downloads include their ground truth.

Download	Images	ZIP size	Use
Lite	60	16.3 MiB	The fixed, outcome-independent set used in the paper's human-guided comparison
Full	2,980	715 MiB	2,500 core plots and 480 matched rerenderings for automatic evaluation

Download and unpack Lite, then load a plot and its reference data. While the repository is private, use hf auth login with an account that has access.

python -m pip install huggingface_hub pandas pyarrow pillow
hf download LinePilot/DigitizerBench lite.zip --repo-type dataset --local-dir digitizerbench
python -m zipfile -e digitizerbench/lite.zip digitizerbench

from pathlib import Path

import pandas as pd
from PIL import Image

root = Path("digitizerbench/lite")
manifest = pd.read_parquet(root / "manifest.parquet")
case = manifest.iloc[0]
image = Image.open(root / case["image_path"])
image_filter = [("image_id", "==", case["image_id"])]
curves = pd.read_parquet(root / "ground_truth.parquet", filters=image_filter)
ticks = pd.read_parquet(root / "tick_ground_truth.parquet", filters=image_filter)

image.show()
for line_id, points in curves.groupby("line_id"):
    print(line_id, points.sort_values("sample_index")[["x", "y"]].head())

For Full, replace lite.zip and digitizerbench/lite with full.zip and digitizerbench/full.

From a plot image to recovered curves

The Altair example below shows two overlapping series from DigitizerBench and the corresponding LinePilot OCR recovery from Figure 1b of the paper.

Benchmark input	LinePilot OCR recovery

_{LinePilot follows each curve across the image and exports its calibrated x, y coordinates.}

Files and ground truth

File	Contents
images/	Input PNGs organized by renderer and plot type
manifest.parquet	One row per image. Join on image_id and resolve image_path relative to the extracted Full or Lite folder. Includes base_id, renderer, plot type, line count, factor settings and seeds.
ground_truth.parquet	One row per curve sample. Group by image_id and line_id, then order by sample_index. Use x and y as the reference in data coordinates. The remaining fields describe the signal, baseline, noise and stacked offsets.
tick_ground_truth.parquet	Visible ticks keyed by image_id, axis and tick_index. Includes the rendered label, numeric value and tick position in top-left-origin image pixels.
factor_levels.json	Factor definitions and levels
design/	Full only. Base configurations, orthogonal-panel assignments, bridge manifest and design summary.

Benchmark design and evaluation

The plots are synthetic. They span single, overlapping and stacked layouts rendered in Matplotlib, Plotly and Altair. Controlled factors include curve shape, baseline, noise, stroke width, color, DPI, blur, figure dimensions, tick styling and axis ranges. Full uses four strength-3 orthogonal-array panels. Each of the 60 Lite configurations has nine renderer × plot-type variants in Full, including its Lite image.

Run your digitizer on the PNGs, then join the recovered data to ground truth by image and match the recovered curves to the reference lines. For the paper's protocol, report failure-penalized capped normalized RMSE (FPC-NRMSE) alongside output success, trusted usability and points per line. FPC-NRMSE assigns unit loss to missing, unusable or catastrophically inaccurate output. See the paper and appendix for the matching and scoring rules.

Lite is nested within Full. For custom train/test partitions, keep all variants sharing a base_id together. Results characterize recovery on these controlled renders. Scanned pages, compression and real-world annotation styles require separate evaluation.

Citation

If you use DigitizerBench, cite LinePilot Digitizer: Line-Plot Recovery with Manual and Automatic Calibration. For the digitizer itself, see the LinePilot releases. Report benchmark or dataset issues on GitHub.

@article{ma2026linepilot,
  title   = {LinePilot Digitizer: Line-Plot Recovery with Manual and Automatic Calibration},
  author  = {Ma, Fengbo and Akhtar, Rayan and Joshi, Aakash H. and Li, Xiaoting and
             Sun, Haijian and Xiang, Zhen and Chen, Xianyan and Zhao, Yiping},
  journal = {arXiv preprint arXiv:2609.19377},
  year    = {2026},
  doi     = {10.48550/arXiv.2609.19377}
}