How Change Detection Works
Summary: Bedrock uses computer vision and AI to compare construction drawings at the pixel level, then categorizes detected changes by type. The process is automatic, requiring no manual alignment or configuration.
Overview
Change detection in Bedrock happens in four stages:
- Ingestion: PDFs are processed into normalized images
- Matching: Sheets are paired between prior and current revisions
- Alignment: Paired sheets are geometrically aligned
- Detection: Differences are identified and categorized
Each stage uses purpose-built algorithms optimized for construction drawings.
Stage 1: Ingestion
When you upload a PDF drawing set, Bedrock:
- Extracts each page as a high-resolution image
- Identifies sheet boundaries and title blocks
- Extracts text layers for metadata
- Normalizes resolution and color profiles
| Parameter | Specification |
|---|---|
| Processing Resolution | 300 DPI equivalent |
| Color Handling | Converted to grayscale for comparison |
| Text Extraction | OCR + native PDF text layers |
Stage 2: Sheet Matching
Bedrock automatically matches corresponding sheets between revisions without requiring manual pairing.
How Matching Works
The system analyzes multiple signals to identify sheet pairs:
| Signal | Weight | Description |
|---|---|---|
| Sheet Number | High | A1.01 in prior matches A1.01 in current |
| Sheet Title | Medium | ”First Floor Plan” matches similar titles |
| Visual Similarity | Medium | Overall drawing structure and content |
| Title Block Position | Low | Location of title block on page |
Handling Edge Cases
| Scenario | How It’s Handled |
|---|---|
| Renumbered sheets | Visual similarity + title matching |
| New sheets | Flagged as “Added” in current revision |
| Deleted sheets | Flagged as “Removed” from prior revision |
| Split sheets | Each new sheet matched to original |
Stage 3: Alignment
Matched sheets must be geometrically aligned before comparison. Traditional tools require manual 3-point alignment. Bedrock does this automatically.
Automatic Alignment Process
- Feature Detection: Identify stable reference points (grid lines, column markers, title blocks)
- Transform Calculation: Compute rotation, scale, and translation
- Image Registration: Apply transformation to align drawings
- Quality Check: Verify alignment meets accuracy threshold
Alignment Accuracy
| Drawing Type | Typical Accuracy |
|---|---|
| Architectural plans | Sub-pixel |
| Structural framing | Sub-pixel |
| MEP layouts | 1-2 pixels |
| Site plans | 2-3 pixels |
Accuracy varies based on drawing quality and the presence of stable reference geometry.
Stage 4: Change Detection
With sheets aligned, the system identifies what changed between revisions.
Detection Methods
| Method | What It Finds |
|---|---|
| Pixel Differencing | Raw visual differences between images |
| Contour Analysis | Changes to lines and shapes |
| Text Comparison | Modified, added, or deleted text |
| Symbol Detection | Changes to common drawing symbols |
Change Categories
Detected changes are automatically categorized:
| Category | Description | Examples |
|---|---|---|
| Addition | New content in current revision | New walls, equipment, notes |
| Deletion | Content removed from prior revision | Removed doors, deleted dimensions |
| Modification | Changed existing content | Moved walls, updated dimensions |
Filtering
Not all differences are meaningful changes. Bedrock filters out:
- Title block updates (dates, revision numbers)
- Watermarks and stamps
- Minor PDF rendering variations
- Compression artifacts
Technical Specifications
| Specification | Value |
|---|---|
| Detection Sensitivity | Configurable (default: medium) |
| Minimum Change Size | 0.1” at print scale |
| Maximum Sheet Size | 36” x 48” (ARCH E) |
| Processing Memory | Optimized for large drawings |
Limitations
Change detection has known limitations:
| Limitation | Impact | Mitigation |
|---|---|---|
| Scanned drawings | Lower alignment accuracy | Use native PDFs when possible |
| Rotated sheets | May require manual verification | System flags suspected rotation |
| Color-coded changes | May miss color-only changes | Grayscale processing |
| 3D views | Less reliable than 2D plans | Focus on plan views |
FAQ
How accurate is change detection?
Graphic element detection exceeds 99% accuracy on native PDFs. Text detection varies based on PDF text layer quality. Scanned drawings have lower accuracy due to image noise.
Does detection work on all drawing types?
Yes, but accuracy is highest on 2D plan views (floor plans, elevations, sections). 3D perspectives and renderings are less reliable due to shading complexity.
Can I adjust detection sensitivity?
Yes. Higher sensitivity catches smaller changes but may increase noise. Lower sensitivity reduces noise but may miss subtle changes. Most users find the default setting optimal.
What causes false positives?
Common causes: PDF regeneration differences, font substitution, compression artifacts, scan quality variations. Filtering algorithms reduce but don’t eliminate false positives.
Key Takeaways
- Change detection happens in four stages: ingestion, matching, alignment, detection
- Sheet matching uses multiple signals (number, title, visual similarity)
- Alignment is automatic with sub-pixel accuracy on most drawings
- Changes are categorized as additions, deletions, or modifications
- Filtering removes noise (title blocks, watermarks, rendering differences)
- Accuracy is highest on native PDFs; scanned drawings have limitations
Last updated: 2026-02-04