jedarden 4409eff058 feat(pdftract-88sk): fix 5x3 table test and add benchmark

Fix the critical 5x3 bordered table test to match acceptance criteria
(5 rows × 3 columns = row_ys.len() == 6, col_xs.len() == 4).

Add missing unit tests:
- test_detect_nested_rectangles: tests handling of nested rectangles
- test_detect_disjoint_tables: tests detection of multiple disjoint tables

Add Criterion benchmark for table detection performance.
Results: ~772 µs for 1000 segments (well under 5 ms requirement).

All 35 table module tests pass.

Acceptance criteria:
- ✅ Detector emits GridCandidate for every closed grid of >= 4 cells
- ✅ Critical test: 5x3 bordered table with row_ys.len()==6, col_xs.len()==4
- ✅ Unit tests: single rectangle, nested rectangles, mixed text+rules, glyph-path noise
- ✅ Public TableDetector::detect_line_based(&PageContext) -> Vec<GridCandidate>
- ✅ Benchmark: < 5 ms on 1000-segment page

Refs: pdftract-88sk, plan section 7.2 line 2571

Co-Authored-By: Claude Code <noreply@anthropic.com>

2026-05-23 21:40:57 -04:00

3.1 KiB

Raw Blame History

Verification Note: pdftract-88sk - Line-based Table Detection

Summary

Implemented line-based table detection for bordered tables. The implementation was already mostly complete in the existing codebase. Fixed the critical 5x3 table test and added missing unit tests (nested rectangles, disjoint tables) plus a benchmark.

Changes Made

Files Modified

crates/pdftract-core/src/table/detector.rs
- Fixed test_detect_5x3_table: Changed from 3 rows × 5 columns to 5 rows × 3 columns to match acceptance criteria (row_ys.len() == 6, col_xs.len() == 4)
- Added test_detect_nested_rectangles: Tests handling of nested rectangles (e.g., table within a table)
- Added test_detect_disjoint_tables: Tests detection of multiple disjoint tables on the same page
crates/pdftract-core/Cargo.toml
- Added criterion = "0.5" to dev-dependencies
- Added [[bench]] section for table_detection benchmark
crates/pdftract-core/benches/table_detection.rs (new file)
- Criterion benchmark testing performance with varying segment counts
- Tests 20, 40, 60, 100, and 1000 segment configurations

Acceptance Criteria Status

Criteria	Status	Notes
Detector emits GridCandidate for every closed grid of >= 4 cells	✅ PASS	`build_grids()` filters by `min_cells` (default 4)
Critical test: 5x3 bordered table returns GridCandidate with row_ys.len()==6, col_xs.len()==4	✅ PASS	Fixed test now correctly draws 5 rows × 3 columns (6 horizontal, 4 vertical lines)
Unit tests: single rectangle	✅ PASS	`test_collect_rectangle`
Unit tests: nested rectangles	✅ PASS	`test_detect_nested_rectangles` (new)
Unit tests: mixed text+rules	✅ PASS	`test_filter_text_object_segments`
Unit tests: glyph-path noise rejected	✅ PASS	`test_filter_text_object_segments`
Public TableDetector::detect_line_based(&PageContext) -> Vec	✅ PASS	Method exists and is public
Benchmark: < 5 ms on 1000-segment page	✅ PASS	Actual: ~772 µs (0.77 ms)

Test Results

test result: ok. 35 passed; 0 failed; 0 ignored

All 35 table module tests pass, including:

Segment creation and manipulation tests
Grid candidate construction tests
Detector tests (segment collection, clustering, intersection finding, grid building)
5x3 bordered table critical test

Benchmark Results

table_detection/dense_table_1000_segments
                        time:   [762.36 µs 772.02 µs 784.69 µs]

Performance is well under the 5 ms requirement for 1000-segment pages.

Implementation Notes

The existing implementation already had:

Segment extraction from PDF path operators (m, l, re, S, s, f, F, B, B*)
Text object filtering (BT..ET) to exclude Type 3 font glyph outlines
Collinear segment clustering with epsilon 1.0 pt tolerance
Gap tolerance of 2.0 pt for merging overlapping collinear segments
Intersection finding between horizontal and vertical segments
Grid construction from intersection points

The main fix was correcting the critical test to match the acceptance criteria (5 rows × 3 columns, not 3 rows × 5 columns).

3.1 KiB Raw Blame History Unescape Escape