Landmarks and Regions: A Robust Approach to Data Extraction
summary
Introduces a landmark-and-region synthesis pipeline that narrows extraction to small ROIs derived from nearby invariant strings, then runs lightweight programs on those regions; uses parent hops and sibling hops for HTML regions and a box summary plus a visual DSL for OCR images. Claims production gains: about 2.95 selectors per program versus 8.51 for prior work, saves ~120 maintenance hours and halves annotation bandwidth in one deployment, and outperforms Azure Form Recognizer with as few as 10 training examples.
AI-written summary · about
chapters (5)
related talks
14mRegion-based Content Enhancement for Efficient Video Analytics at the Edge
01h 00mOpening Keynote
26mAdvanced Document Parsing for AI Workflows
22mOCR, Recognition and Information Extraction from Images
31mHow our neural networks find polygons to understand your documents better
41mKEYNOTE: Image processing, AI, and autonomous systems
41mOCR in the Wild World of Social Media
46mML Use Cases at Think Big Analytics