Code Capture AI — Computer Vision Pipeline

AI-powered ADA/AODA building code compliance analysis for architectural floor plans. Upload a floor plan PDF, get back accessibility violations detected by computer vision.

Built by Geopogo (Berkeley, CA) in collaboration with Architecttura.

How It Works

PDF floor plans
  → Image extraction (PyMuPDF @ 300 DPI)
  → Auto-annotation (Grounding DINO / Gemini 2.5 Flash)
  → Manual review & correction (Label Studio)
  → Train/val split (80/20)
  → YOLOv12 training (ultralytics)
  → Trained model serves predictions via API

Detection Classes

The model detects 10 accessibility-relevant architectural elements:

ID	Class	ADA Relevance
0	door	Clear width 32", maneuvering clearance, hardware
1	window	Operable hardware in accessible units
2	corridor	Min 36" width, slope limits
3	toilet	Centerline 16-18" from wall, clear floor space
4	stairs	Tread/riser dims, handrails; flags need for accessible alternative
5	ramp	Max 1:12 slope, 36" width, handrails, landings
6	elevator	Min 51"x68" car, Braille controls
7	dimensions	Dimension lines — critical for verifying all ADA clearances
8	room_tag	Identifies room function → applicable ADA requirements
9	sink	Knee clearance 27", rim height 34" max

Quick Start

# 1. Setup
python -m venv .venv
source .venv/bin/activate
pip install -r requirements.txt

# 2. Extract floor plan images from PDFs
python pdftoimg.py

# 3. Auto-annotate images (choose a backend)
python auto_annotate.py --model grounding-dino       # local, zero-shot
python auto_annotate.py --model gemini --dry-run      # requires GOOGLE_API_KEY

# 4. Import pre-annotations into Label Studio for review
python import_to_labelstudio.py

# 5. Run the full pipeline end-to-end
python run_pipeline.py

# 6. Or run individual steps
python run_pipeline.py --skip-extract --skip-annotate  # resume from split
python run_pipeline.py --review                        # pause after annotation

Pipeline Scripts

Script	Purpose
pdftoimg.py	Extract all pages from PDFs as PNGs at 300 DPI
auto_annotate.py	Zero-shot annotation via Grounding DINO or Gemini 2.5 Flash
import_to_labelstudio.py	Convert YOLO labels to Label Studio JSON for manual review
split_dataset.py	Shuffle and split labeled data into train/val (80/20, seed=42)
train.py	YOLOv12 training with augmentation tuned for B&W line drawings
run_pipeline.py	Orchestrate all steps with skip flags and a review gate

Project Structure

CodeComplianceAI-CV/
├── auto_annotate.py          # Auto-labeling (Grounding DINO / Gemini)
├── import_to_labelstudio.py  # YOLO → Label Studio JSON converter
├── pdftoimg.py               # PDF page extraction
├── split_dataset.py          # Train/val split
├── train.py                  # YOLOv12 training config
├── run_pipeline.py           # End-to-end orchestrator
├── test_auto_annotate.py     # Tests
├── dataset.yaml              # YOLO dataset config (10 classes)
├── requirements.txt          # Python dependencies
├── images/                   # 117 source floor plan images
├── image_chunks/             # Images split across 6 annotators
├── docs/                     # Setup guides (Label Studio ML backend)
└── dataset/                  # Generated train/val splits (gitignored)
    ├── images/{train,val}/
    └── labels/{train,val}/

Training Configuration

The training script (train.py) is configured for architectural line drawings:

• Model: YOLOv12n (nano) with COCO pretrained weights
• Image size: 1024px (source images are 10800x7200)
• Color augmentation: Disabled (hsv_h/s/v = 0) — floor plans are B&W
• Geometric augmentation: 90° rotation, horizontal/vertical flips
• Mosaic: Disabled in final 10 epochs for fine-tuning

Auto-Annotation Backends

Grounding DINO (default)

Zero-shot object detection via HuggingFace Transformers. Runs locally, no API key needed. Supports MPS (Apple Silicon), CUDA, and CPU.

python auto_annotate.py --model grounding-dino --threshold 0.25

Gemini 2.5 Flash

Google's multimodal model with structured JSON output. Requires GOOGLE_API_KEY env var.

export GOOGLE_API_KEY=your_key_here
python auto_annotate.py --model gemini

Dataset

• 117 floor plan images (10800x7200 and 10800x9000 resolution)
• 33 real-world architectural PDFs from Architecttura (Floorplan_Dataset3_3_10_2026/)
• 6 annotators labeling in parallel via Label Studio
• Export format: YOLO (.txt label files with normalized bounding boxes)

Tech Stack

• Python 3.11+
• YOLOv12 (ultralytics) — attention-centric object detection
• Label Studio — annotation and review
• Grounding DINO — zero-shot pre-annotation
• Gemini 2.5 Flash — multimodal pre-annotation
• PyMuPDF (fitz) — PDF to image extraction
• Pillow / OpenCV — image processing

Tests

pytest

License

Proprietary — Geopogo, Inc.