Turning table extraction into an object-detection problem

What it does TATR is an object-detection model—built on DETR—that finds tables in document images and breaks them down into rows, columns, and cells. It ships with pre-trained weights, an inference pipeline that emits HTML or CSV, and the official PubTables-1M dataset used to train it. The repo also implements GriTS, the grid-based metric the authors proposed to evaluate table structure recognition.

The interesting bit Instead of treating table extraction as a sequence or layout problem, it frames everything as bounding-box detection: tables, rows, columns, and even projected headers are just objects to localize. That architectural choice means you can train it like a standard vision model, though you still need to bring your own text extraction—TATR detects structure from pixels, but OCR or PDF text is required to fill in the actual cell content.

Key highlights

• Pre-trained models include TATR-v1.1 variants trained on PubTables-1M, FinTabNet.c, or both combined.
• PubTables-1M contains nearly 950k fully annotated tables with bounding boxes in both image and PDF coordinates.
• The repo provides training scripts, evaluation code, and canonicalization utilities for existing datasets like FinTabNet and SciTSR.
• GriTS metrics are included for measuring topology, content, and location similarity independently.
• Detection AP50 on PubTables-1M is reported at 0.995; structure AP50 at 0.970 for TATR-v1.0.

Caveats

• The model detects structure from pixels; populating cells with text requires a separate OCR or PDF text extraction step.
• Training your own model is expected for domains outside the provided PubTables-1M and FinTabNet.c datasets.

Verdict Researchers and engineers building document-parsing pipelines should look here for a well-documented baseline and a massive labeled dataset. If you need an end-to-end, OCR-inclusive one-liner, this is only half the pipeline.