Digital twin
A photoreal, navigable 3D model of the whole property. Pan, zoom and walk through it in a browser, no app install.
Property condition · AI inspection
Three cameras. One 3D model. Every defect flagged by AI.
A drone flies the roof, Meta Ray-Ban glasses walk the inside, a phone catches the detail. We fuse all three into a single 3D model, then AI flags every crack, stain, spall and missing tile in place.
Live 3D · drag to rotate · scroll to zoom
What you get
One capture. Three deliverables.
From a single fused capture, the AI produces a navigable 3D model, every defect flagged on it, and a shareable condition report.
Defect detection
Every crack, spall, stain and missing tile auto-flagged, labeled and pinned to its exact spot on the model.
Condition report
A shareable report: each defect with its photo, location, severity and AI confidence. Ready for your file, your client or your insurer.
Inside the home
The walk-through, in 3D.
The drone covers the outside. A smart-glasses or smartphone pass captures every room. Same 3D model. Same defect detection. Same shareable link.
Live 3D interior · drag to rotate · scroll to zoom
Three sources, one model
Each camera sees what the others can't.
Drone, Meta Ray-Ban glasses and phone aren't alternatives; they're layers. Together they cover roof to subfloor, then fuse into one navigable 3D model with georeferenced defect markers.
Drone
Best for roof + upper facade. A 30-min flight at three altitudes produces ~1,000 RGB frames.
- DJI Mavic 3 / Mini 4 Pro / Skydio X10
- 3-altitude protocol: high nadir + oblique orbits + facade orbit
- GPS-anchored scale + georeference
Smart glasses
Ego-perspective ground-level capture. Hands-free walkthrough of interior + close-up details.
- Meta Ray-Ban / Oakley Sphaera / Aria 2
- Ideal for interior + tight passages
- Low-friction — walk and capture
Smartphone
Lowest barrier. iPhone Pro's LiDAR sensor adds depth priors; mid-range Android still works.
- iPhone 12 Pro+ ideal · most Android 8GB+
- 1080p30 video, manual exposure
- Closeup defect capture
The technology
From three cameras to one measurable model.
Drone, Meta Ray-Ban glasses and a phone each capture what the others miss. The pipeline fuses them into a single navigable 3D model with every defect pinned in place.
01
Capture
Drone · Meta Ray-Ban · phone
Three cameras, three zones. The drone flies the roof and upper facade; Meta Ray-Ban glasses walk the interior and tight spaces hands-free; a phone grabs close-ups of anything that needs detail. One property, one session.
⏱ ~30 min on-site
02
Register & fuse
COLMAP structure-from-motion
The step that makes it one model. Structure-from-motion solves every frame from all three devices into a single shared coordinate frame, so aerial, ego-level and close-up imagery line up to the millimetre.
The hard part, and our moat
03
Reconstruct
Gaussian Splatting (gsplat) + PGSR mesh
A photoreal 3D Gaussian Splat you can navigate in a browser, plus a PGSR mesh for measurement-grade geometry: 5 to 15 mm wall error against LiDAR ground truth.
⏱ ~45 min compute
04
Detect
YOLOv11-seg + SAM 2
Defects are segmented on the flat source frames, then lifted into the 3D model through the shared camera geometry from step 2, so every crack, spall and stain lands at its exact position on the wall.
05
Deliver
Interactive 3D + annotated report
A browser link: pan, zoom, and click any marker for the source photo, AI confidence and note. Re-capture at 6 / 12 / 24 months for an automatic "what changed" diff on the same 3D scaffold.
⏱ 24 hr turnaround
What we detect
Seven defect classes today. Growing.
Trained on CUBIT-InSeg, SDNET2018 and AU-specific building datasets. Each class has a published accuracy benchmark.
Cracks88% F1
Concrete, masonry, render. Measured to nearest 2 mm.
Spalling82% F1
Surface loss on concrete and brickwork.
Corrosion79% F1
Steel, exposed rebar, gutters, fasteners.
Missing tiles75% F1
Roof, facade, floor — anywhere with a uniform pattern.
Water damage74% F1
Staining, efflorescence, soft-edge moisture marks.
Render failure71% F1
Bubbling paint, blown render, peeling coatings.
Vegetation overhang85% F1
Trees and growth contacting the building envelope.
Inside the model
Pixel-level segmentation, built for real walls.
A condition report needs more than a box around a problem. Our model outputs a pixel-accurate mask for every defect, so cracks and spalling are measured, not just spotted.
Detection runs in two stages: a segmentation network localizes each defect, then a promptable segmenter (SAM 2) traces it to a clean per-pixel mask. From those masks we read a crack's width and a spall's area as real numbers, in mm and mm².
The hard part is never a clean lab crack. The model is trained and scored on the three conditions that break naive detectors:
- High local contrast under harsh light and deep shadow, where shadow lines mimic real cracks.
- Noisy backgrounds: vegetation, cabling, air-con units and fixtures crowding the target.
- Low-texture surfaces, where a hairline crack is almost invisible to the eye.
Why SafeDetect
What we do that nobody else does.
SafeDetectDrone-only inspectionAerial monitoring
Cross-source capture (drone + glasses + phone)All threeDrone onlyAerial only
Interior + close-up coverageYesPartialNo
3D Gaussian Splat outputPhotoreal3D twin2D ortho
Open-source stack — no vendor lockYesProprietaryProprietary
AU-native operator networkBuildingIn-houseOwn fleet
Time-series re-inspection at 6 / 12 / 24 monthsYesPartialYes
14,000+
Defects classified
on CUBIT-InSeg + SDNET2018 training data
30 min
On-site capture
typical drone flight per 2-3 storey home
24 hrs
Report delivered
from capture upload to interactive 3D link
A real example
What an inspector sees vs what we surface.
Take a 1960s weatherboard cottage in Bracken Ridge. Here's the difference between a 50-photo PDF and a navigable 3D model with detection.
50-photo PDF report
- 2-3 hours on-site walk-through
- Inspector takes ~50 photos, captions each by hand
- 20-30 page PDF delivered 2-5 days later
- Buyer can't relate "crack on north wall, photo 23" to the actual location
- No comparison if you re-inspect 12 months later — different photos, different angles
Navigable 3D model + AI markers
- 30-min drone flight + optional ground walkthrough
- 2.4M-point Gaussian Splat — every facade pixel captured
- Browser link delivered within 24 hours of capture
- 14 defects auto-flagged at exact 3D positions — click each for photo + note
- Re-inspect at 6 / 12 / 24 months: same model, "what changed" diff
Trust
Built to inspection-grade standards.
CASA-certified operators
Every flight in the SafeDetect network is conducted by a licensed AU operator with current ReOC + public-liability cover. We verify both on onboarding.
Open-source pipeline
COLMAP, gsplat, YOLOv11-seg, SAM 2 — every component in the processing stack is Apache-2 / MIT / BSD licensed. No black box. No vendor lock.
AU-trained models
Defect detection fine-tuned on AU building stock — weatherboard, brick veneer, render, Colorbond, tiled roof. Trained for what you actually have.
Privacy-first
Faces, plates, neighbouring property auto-masked before processing. Source footage retained only for the inspection window, then purged.
Measurement-grade geometry
PGSR mesh extraction gives 5-15 mm wall-error vs LiDAR ground truth. Good enough for a crack width or a spall area to be quoted in mm².
Auditable + repeatable
Capture, processing parameters and detection thresholds are all logged. Two inspections of the same property produce comparable, defensible numbers.
SafeDetect is a condition-detection product, not a substitute for a licensed building inspection. Always commission an AS 4349.1-certified inspector before contract exchange.
FAQ
Common questions.
Is this a replacement for a licensed building inspection?
No. SafeDetect surfaces visible exterior + interior defects from imagery — what an inspector would see with eyes, just faster and georeferenced. We do not replace structural assessment, pest inspection, plumbing or electrical certification. Always commission an AS 4349.1-certified inspector before contract exchange.
How accurate is the defect detection?
Published benchmarks place facade-crack detection at ~88% [email protected], spalling at ~82%, corrosion at ~79%. Accuracy varies by lighting, capture angle and material. We surface confidence scores per detection so you can see when we are guessing. False negatives are possible — the operator and inspector review the model alongside the auto-flags.
What if I already have drone footage?
Upload it. The processing pipeline accepts any RGB photo set or video from any drone — DJI, Skydio, Autel, custom. We extract camera poses via COLMAP, train the Gaussian Splat, and run defect detection on the source frames. No re-flight needed.
How do I view the 3D model?
A browser link, mobile or desktop, works in Chrome / Safari / Firefox. Drag to rotate, scroll to zoom, click any defect marker for the source photo + AI confidence + your note. Shareable with your conveyancer, partner, builder — no app install.
What if the property has glass facades or reflective surfaces?
Pure Gaussian Splat struggles with glass — view-dependent reflections create "ghost" Gaussians. We cross-check with a photogrammetric mesh for those regions and document the limit honestly in the report. Most AU residential is brick/weatherboard/render where this is not an issue.
Can I re-inspect the same property later?
Yes — that is one of our most valuable features. We capture the same property at 6 / 12 / 24 months and produce a "what changed" comparison: new defects, defects that worsened, defects that resolved. Same 3D scaffold; differential overlay.
Who owns the 3D model + footage?
You do. The customer who paid for the inspection owns the captured footage and the resulting 3D model. SafeDetect keeps a backup for 12 months for re-inspection eligibility, after which it is purged unless you opt in to retention.
Is the operator network covered by insurance?
Yes — public-liability insurance is mandatory for every operator on the network, minimum $20M. We verify the certificate on onboarding and re-check at renewal. Any claim arising from a flight is handled by the operator's insurer; SafeDetect is the platform, not the flier.
Get started
Want to see what your building looks like in 3D?
Book a capture on one property. We'll walk you through the 3D model and the defects we find, no commitment.
Or email us at [email protected]