# SignverOD: Signature Object Detection in Scanned Documents

> A CC0-licensed dataset of 2,576 scanned document images with 7,103 bounding box annotations for detecting signatures, initials, redactions, and dates.

- Authors: Bevaya AI Team
- Published: 2026-03-13
- Updated: 2026-03-15
- Dataset: SignverOD <https://www.kaggle.com/datasets/victordibia/signverod>
- Code: SignVer (Signature Verification Library) <https://github.com/victordibia/signver>
- Source: Bevaya Labs, <https://labs.bevaya.ai/>

---


## TL;DR

SignverOD provides **2,576 scanned document images** with **7,103 bounding box annotations** across four classes (Signature, Initials, Redaction, Date). It targets a narrow but practically important slice of document layout: detecting handwritten artifacts and intentional occlusions in business documents. Licensed CC0 (public domain).

---

## What kind of paper is this?

**Dominant:** $\Psi\_{\text{Resource}}$

The contribution is the dataset itself. There is no associated peer-reviewed publication; the dataset was released on Kaggle with a self-published technical report (now offline) and a companion signature verification library (SignVer).

---

## What is the motivation?

Detecting the presence and location of handwritten artifacts (signatures, initials, dates) in scanned offline documents supports multiple downstream tasks: signature verification, document tagging, and categorization. Most existing layout analysis datasets either omit these classes entirely or include them as minor additions (e.g., IIIT-AR-13K's Signature class accounts for fewer than 600 of its 23,000 annotations). SignverOD focuses exclusively on this subset of form-level elements, providing denser supervision for these specific classes.

The dataset also includes **Redaction** as a class, which is unusual. In the Matter vs. Meaning framework, redaction maps to a distinct visual primitive (`Redaction`) that requires its own processing logic (masking/logging rather than OCR).

---

## What is the novelty?

The novelty is practical rather than methodological: a focused, freely licensed (CC0) collection of diverse business documents annotated specifically for form-element detection. The four classes map cleanly to the Matter vs. Meaning taxonomy:

| SignverOD Class | Matter Primitive | Meaning Role | Processing Logic |
| :--- | :--- | :--- | :--- |
| **Signature** | `Field` / `Text` | `Signature` | Verification / Matching |
| **Initials** | `Field` / `Text` | `Signature` (variant) | Verification / Matching |
| **Redaction** | `Redaction` | (none; primitive is self-describing) | Masking / Log |
| **Date** | `Text` (`is\_handwritten: True`) | `Value` | HTR |

The source documents span memos, emails, bank cheques, lease agreements, letters, and invoices, providing more domain diversity than typical scientific-paper layout datasets.

---

## What experiments were performed?

No experiments are reported in the dataset release. The companion SignVer library uses the dataset's Signature annotations for a detection module (bounding box localization), followed by cleaning and metric learning for verification, but no benchmark results on the detection task are published.

---

## What are the outcomes/conclusions?

As a dataset-only release without a formal evaluation, there are no reported metrics. The dataset's value is as a training resource for the specific classes it covers.

---

## Reproducibility

### Models

No models released.

### Algorithms

Not applicable (dataset release only).

### Data

- **Images:** 2,576 scanned document pages.
- **Annotations:** 7,103 bounding boxes across 4 classes (Signature, Initials, Redaction, Date).
- **Document types:** Memos, emails, bank cheques, lease agreements, letters, invoices.
- **Image sources:**
  - [Tobacco800](https://www.kaggle.com/sprytte/tobacco-800-dataset): scanned tobacco litigation documents.
  - [NIST Special Database 2](https://www.nist.gov/srd/nist-special-database-2): structured tax forms.
  - [Bank Cheques Dataset](https://www.kaggle.com/saifkhichi96/bank-checks-signatures-segmentation-dataset): cheque images with signatures.
  - [GSA.gov Lease Documents](https://www.gsa.gov/real-estate/real-estate-services/leasing/executed-lease-documents): U.S. government lease agreements.
- **License:** CC0-1.0 (public domain). Source collections are public records (Tobacco800 from tobacco litigation discovery; NIST SD-2 is a U.S. government product; GSA leases are government documents). The bank cheques dataset (BCSD) is listed as CC0 on Kaggle. Tobacco800 has no formal license declaration, though its underlying documents are public legal records.
- **Format:** Bounding box annotations (object detection format). Available on Kaggle (~760 MB compressed).
- **Splits:** Not documented in the Kaggle description. Users should verify whether predefined train/val/test splits are included.

### Evaluation

No formal evaluation benchmarks or baselines reported.

### Hardware

Not applicable.

---

## BibTeX

```bibtex
@article{Dibia2022signverod,
  author = {Dibia, Victor},
  title = {A Dataset for Handwritten Signature Object Detection in Scanned Documents},
  year = {2022},
  publisher = {victordibia.com},
  journal = {victordibia.com},
  url = {https://www.kaggle.com/datasets/victordibia/signverod}
}
```