Awesome Table Detection

A practical guide to table detection — finding where tables are in documents and images.

Whether you're building a document processing pipeline, extracting data from PDFs, or working with photos of tables, this repo helps you understand the landscape and pick the right tools.

---

What is table detection?

Table detection answers one question: "Where are the tables on this page?"

Input	Output
A document page (PDF, scan, photo)	Bounding boxes or polygons around each table

That's it. Table detection finds the regions — it doesn't extract the data inside (that's table structure recognition + OCR, which comes after).

Why it matters

Tables are everywhere: invoices, reports, scientific papers, forms, receipts. Before you can extract data from a table, you need to know where it is. Good detection is the foundation of any table extraction pipeline.

The two worlds of tables

Document tables	Tables in the wild
PDFs, scans, reports, papers	Photos, screenshots, whiteboards
Usually axis-aligned	Often rotated, curved, perspective-distorted
Consistent layouts	Varied lighting, occlusions, noise
Easier to detect	Harder to detect

---

Quick start: Which solution should I use?

flowchart TD
    A[What's your input?] --> B{PDF with<br>selectable text?}
    B -->|Yes| C[Try heuristic tools first]
    C --> C1[Tabula / Camelot]
    C1 --> C2{Works well?}
    C2 -->|Yes| DONE[Done!]
    C2 -->|No| D

    B -->|No| D{Scans or<br>document images?}
    D -->|Yes| E{Want to train<br>your own?}

    E -->|No| E1[Use pretrained models]
    E1 --> E1a[Table Transformer]
    E1 --> E1b[Docling]
    E1 --> E1c[Surya]

    E -->|Yes| E2[Train custom detector]
    E2 --> E2a[YOLO + DocLayNet/PubTables]
    E2 --> E2b[Detectron2 / MMDetection]
    E2 --> E2c[Fine-tune TATR]

    D -->|No| F{Photos or<br>screenshots?}
    F -->|Yes| G[Use wild-trained models]
    G --> G1[Fine-tune on TabRecSet/WTW]
    G --> G2[Consider polygon detectors]

    A --> H{No labeled<br>data?}
    H -->|Yes| I[Bootstrap with VLMs]
    I --> I1[Grounding DINO]
    I --> I2[Florence-2 Table Detection]
    I --> I3[Then train a dedicated detector]

Loading

Your situation	Recommended approach	Tools
Clean PDFs (selectable text)	Heuristic extraction	Tabula, Camelot
Scans / images (use pretrained)	Ready-to-use models	Table Transformer, Docling, Surya
Scans / images (train your own)	Custom training	YOLO, Detectron2, MMDetection
Photos / wild tables	Domain-specific models	Train on TabRecSet or WTW
No labeled data	VLM bootstrapping	Grounding DINO, Florence-2

For detailed guidance, see Choosing a solution

---

Key datasets

Document tables

Dataset	Size	Best for	Link
PubTables-1M	575k pages	Scientific PDFs	HuggingFace
DocLayNet	80k pages	Business documents	GitHub
PubLayNet	360k pages	Layout detection	GitHub
TableBank	417k tables	Pretraining	GitHub

Wild / camera tables

Dataset	Size	Best for	Link
TabRecSet	32k images	Real-world scans	GitHub
WTW	14k images	Distortions, occlusions	GitHub

Diverse document tables

Dataset	Size	Best for	Link
STDW	7k images	Diverse document layouts	GitHub

Full dataset catalog: Datasets documentation

---

Popular models

Model	Type	Speed	Best for	Link
Table Transformer (TATR)	DETR	Medium	Document tables	HuggingFace
YOLO fine-tunes	One-stage	Fast	Production, high volume	Ultralytics
CascadeTabNet	Mask R-CNN	Slower	High accuracy	GitHub
Grounding DINO	Open-vocab	Medium	Zero-shot, weak labeling	GitHub

Full model guide: Model families

---

Document processing pipelines

If you want a complete solution (not just detection):

Tool	What it does	Link
Docling	PDF → structured data (tables, text, figures)	GitHub
Unstructured	Document ETL for LLMs	GitHub
Marker	PDF → Markdown	GitHub
deepdoctection	Configurable DocAI pipelines	GitHub

---

Repo structure

docs/                      # Start here — narrative guides
├── 00-scope-and-definitions.md   # What is table detection?
├── 01-history-timeline.md        # How we got here
├── 02-problem-decomposition.md   # Breaking down the problem
├── 03-datasets.md                # All the datasets
├── 04-model-families.md          # Types of detectors
├── 05-evaluation-and-benchmarks.md  # How to measure success
├── 06-open-source-solutions.md   # Free tools
├── 07-commercial-solutions.md    # Cloud APIs
├── 08-vlms-and-weak-labeling.md  # Using VLMs for detection
└── 09-choosing-a-solution.md     # Decision guide

model-zoo/                 # Curated model cards
├── document-layout-detectors.md
├── table-only-detectors.md
├── vlm-grounded-detectors.md
└── yolo-recipes.md

datasets/                  # Dataset quick-reference
├── documents.md           # PDF/scan datasets
└── in-the-wild.md         # Photo/camera datasets

evaluation/                # Metrics and protocols
├── metrics.md             # AP, IoU, F1 explained
└── protocols.md           # Standardized evaluation

resources/                 # Links and references
├── papers-and-surveys.md  # Key papers
├── github-and-hf-links.md # Quick links
└── communities.md         # Where to get help

---

How to navigate this repo

If you want to...	Go to...
Understand the basics	Scope and definitions
Pick a solution quickly	Choosing a solution
Find a dataset	Datasets
Compare models	Model families
Evaluate fairly	Evaluation
Use VLMs for labeling	VLMs and weak labeling
Find papers	Papers and surveys
Get quick links	GitHub and HF links

---

Key takeaways

1. Table detection ≠ table extraction — detection finds regions, extraction gets the data
2. Domain matters more than architecture — a model trained on your document type beats a "better" model trained on different data
3. Start simple — heuristic tools work great on clean PDFs; don't over-engineer
4. Preprocessing is critical — PDF rasterization DPI, resize policy, and augmentations often matter more than model choice
5. VLMs are enablers, not replacements — use them to bootstrap labels, not as your production detector

---

Contributing

Found a missing dataset? Know a better model? See CONTRIBUTING.md.

---

License

See LICENSE.