TFLOP: Filtered PubTabNet Images and OCR Results for Table Recognition

Upstage's TFLOP dataset is a processed version of PubTabNet, curated for table structure recognition research. It contains a filtered subset of document images with erroneous samples removed and includes corresponding Optical Character Recognition inference results. The dataset was last updated on January 23, 2025.

Use Cases

• Training table structure recognition models based on filtered document images.
• Benchmarking OCR-assisted layout analysis pipelines based on the included OCR inference results.
• Developing or evaluating layout pointer mechanisms for document understanding, as referenced in the associated paper.

Strengths

• Dataset is derived from the established PubTabNet benchmark.
• Includes pre-processed OCR outputs corresponding to the filtered images.
• Erroneous samples from the original source have been removed.

Limitations

• Description metadata is limited; actual data quality requires manual inspection after download.
• Column-level documentation is absent; field semantics must be inferred after download.
• Row count is unknown, which may limit suitability assessment.

Provenance

Source

PubTabNet dataset

Collection Method

Filtered subset of images with erroneous samples removed; includes OCR inference results.

Freshness

Last updated 2025-01-23 04:27:54; freshness should be verified.