Bounding Box

A bounding box is the rectangle — typically four coordinates: left, top, width, height (or two corner points) — that locates an element within a document image: a word, a text line, a field value, a table, a signature, a stamp. It is the fundamental unit of spatial information in document AI, appearing everywhere: OCR engines emit a box per recognized word; layout models emit boxes per region; extraction systems attach a box to each field value; and annotation for training consists largely of humans drawing them. Coordinates may be absolute pixels or normalized to page dimensions, the latter surviving resizing and rendering differences.

Boxes do two jobs beyond mere location. As model inputs, they carry the spatial signals that document understanding depends on — layout-aware models consume each token's box alongside its text, which is how they learn that a number's meaning depends on which column header sits above it. As provenance, they anchor outputs to evidence: an extracted "Total: 4,860.00" with a bounding box is a claim you can verify by looking exactly where the system looked. That anchoring powers the review interfaces where humans confirm low-confidence values in seconds, the audit trails that trace database records back to source pixels, and the highlighted citations in document Q&A.

The rectangle is an approximation, and its limits matter at the edges: rotated or curved text overflows an axis-aligned box (quadrilaterals or oriented boxes handle this), tightly packed content produces overlapping boxes, and a box says where but not what — pairing geometry with semantics is the rest of the pipeline's job. Practical systems also navigate coordinate-space pitfalls: PDF points versus rendered pixels, page rotation metadata, and resolution mismatches between the image a model saw and the one a viewer displays, any of which can silently shift a "verified" highlight onto the wrong text.