Back to Compare
Overlay Compare Engine

Visual AI Comparison

Transform drawing comparison from tedious manual overlay to instant visual intelligence. Our cascade alignment system achieves 99%+ accuracy across scale variations, rotations, and different scan qualities.

99%+ Alignment Success
~5s Visual Overlay
4-Level Cascade

The Overlay Pipeline

Five stages transform two document revisions into an intelligent visual comparison.

Fuzzy Matching

Intelligent pairing of sheets across revisions

Click "Visualize Pipeline" to see the process
98%96%94%
A-101
Architectural
A-102
Architectural
M-001
Mechanical
Original
4-A-101
Architectural
4-A-102
Architectural
4-M-001
Mechanical
Revised
Fuzzy matching: Levenshtein + Pattern
3 sheets matched

Fuzzy Sheet Matching

Real-world documents use inconsistent naming. Our normalization algorithm handles revision prefixes, OCR errors, and naming variations automatically.

Document A (Old)

A-101
A-102
MH1O-A
M-001
A-103Deleted

Document B (New)

4-A-101
4-A-102
4-MH10-A
4-M-001
4-A-104New

Fuzzy Matching Rules

Revision Prefix

4-A-101 → A-101

OCR Error (O → 0)

MH1O-A → MH10-A

Case Matching

a-101 → A-101

Cascade Alignment Strategy

Four alignment methods tried in sequence, each with increasing robustness. The system stops when confidence exceeds 88%.

METHOD 1

Scale-Invariant Features

Robust feature detection across scales and rotations

0%

confidence

Threshold: 88%
METHOD 2

Accelerated KAZE

Fast nonlinear scale space for efficiency

0%

confidence

Threshold: 88%
METHOD 3

Oriented FAST/BRIEF

Ultra-fast binary descriptor matching

0%

confidence

Threshold: 88%
METHOD 4

Phase Correlation

Frequency-domain alignment with grid search

0%

confidence

Threshold: 88%

Drawing-Focused Fallback

When initial alignment confidence is low (<88%), the system applies drawing masks to focus feature detection on actual CAD content, avoiding misleading matches from title blocks and sheet numbers.

Change Detection

Red/blue pixel-level comparison with connected components analysis groups nearby changes into meaningful regions.

101ELECNEWMECHOFFICECONFSTORAGER1R2R3
Removed
Added
Modified
R1removal

Wall Removed

R2addition

Room Added

R3mixed

Modified

Clip-Based AI Analysis

Instead of analyzing the full image, we extract focused clips around detected regions. This grounds the AI analysis and enables artifact filtering.

AI Analysis Results
R1
high

Load-bearing wall removed between rooms 101 and 102. Requires structural review.

R2
medium

New storage room added with dimensions 12' x 15'. Includes electrical outlet.

R3
Artifact

Minor drafting artifact from scan alignment. Not a real change.

Two-Phase Architecture

Visual overlay returns instantly while AI analysis runs asynchronously, unblocking the user interface immediately.

Phase 1: Visual Overlay

~2-5 seconds

Lambda function handles image alignment, generates red/blue overlay, and extracts modification regions.

Phase 2: AI Analysis

~10-30 seconds (async)

Vision LLM analyzes extracted clips, provides semantic descriptions, filters artifacts, and assesses severity.

Why Two Phases?

Instant Feedback

Visual overlay ready in seconds

Non-Blocking

AI runs asynchronously

Progressive Enhancement

Insights arrive as ready

99%+

Alignment Success

~5s

Visual Overlay

4

Cascade Methods

88%

Confidence Threshold