Module visual 

Visual NER evaluation metrics.

Overview

Visual NER (VisualNER) extracts entities from images and documents, producing entities with both text spans and bounding box locations.

Evaluation Metrics

From VisualNER benchmarks (FUNSD, SROIE, CORD):

Metric	Description
Text F1	Standard NER F1 on extracted text
Box IoU	Intersection-over-Union of bounding boxes
End-to-End F1	Correct text AND box (>50% IoU)

Benchmark Datasets

Common benchmarks include FUNSD, SROIE, CORD, and DocVQA.

Research Alignment

LayoutLMv3 (arXiv:2204.08387) is a representative reference for multimodal document understanding (text + layout + pixels).

Structs

BoundingBox

A bounding box in normalized coordinates (0.0-1.0).

VisualEvalConfig

Configuration for visual NER evaluation.

VisualGold

A gold standard visual entity with text and bounding box.

VisualNERMetrics

Visual NER evaluation metrics.

VisualPrediction

A predicted visual entity.

VisualTypeMetrics

Per-type visual metrics.

Functions

evaluate_visual_ner

Evaluate visual NER predictions against gold standard.

synthetic_visual_examples

Generate synthetic visual NER examples for testing.