POTATR: Innovative Lightweight Model Achieves Superior Page-Level Table Extraction

Researchers have developed the Page-Object Table Transformer (POTATR), a new lightweight image-to-graph model designed for accurate and efficient table extraction from large-scale documents. With only 29 million parameters, POTATR significantly improves upon existing models by achieving a GriTSCon score of 0.964 on the PubTables-v2 Single Pages benchmark, outperforming larger models while being over 130 times faster and approximately 300 times cheaper. The model's output is spatially grounded, allowing for visual verification and geometric text assignment, and it can be integrated with external OCR for scanned documents and techniques like cross-page merging for full-document table extraction. Code and models will be made available upon release.