Control-room fusion
One operator view instead of a 48-tile wall — tracks carry a single ID across every camera on site.
Tiles replaced 48 → 1
Operators per shift 6 → 2
Real-time sensor fusion API
Every camera on site — one live 3D scene.
HyperFUSE ingests unsynchronized 2D streams and returns a metrically accurate, temporally aligned 3D world model in 11 ms end to end. One API call — no rig, no calibration targets, no offline pass.
Free through 50,000 frames — keys issue instantly, no card.
Scene manifest
Cameras bound 8
Tracks in scene 42
Mean geometric error 1.9 cm
Calibration pass Self-solved
End-to-end latency 11 ms
Sampled live from the junction reference deployment — 8 cameras, 4K, mixed lighting.
veh-04 · 12.4 m
ped-19 · 8.1 m
ped-22 · 9.6 m
veh-07 · 21.8 m
cyc-03 · 18.0 m
ped-31 · 15.2 m
veh-11 · 28.5 m
ped-44 · 6.3 m
Solver
Extrinsics Locked
Reprojection 0.42 px
Clock skew ±2 ms
Solve age 4 h 12 m
Bound 8/8
cam-01 4K·30
cam-02 4K·30
cam-03 1080·25
cam-04 4K·30
cam-05 4K·30
cam-06 1080·30
cam-07 4K·25
cam-08 4K·30
3D view
Watertight mesh plus world-space tracks. Query by volume, not by camera — "who is inside bay 4" is one call.
Tracks 42
Vertices 1.2 M
Drift 0.4 cm
Latency 11 ms
Uptime 99.98%
Developer platform
Three lines from stream to scene.
Bind camera IDs, read fused frames. Extrinsics self-solve on the first 200 frames — nothing to measure, nothing to print, nothing to hold up in front of a lens.
Point at the streams
RTSP, WebRTC, or a LENS NVR group. Mixed makes, mixed resolutions, no shared clock.
Let extrinsics solve
Relative pose converges on the first 200 frames from scene content alone — no boards, no surveyed points.
Read fused frames
World-space tracks, watertight mesh, per-view depth. Same object IDs across every camera.
Get API key Unity · Unreal Engine · WebXR · ARKit · Agora
from hyperfuse import Scene scene = Scene.attach(["cam-01", "cam-02", "cam-04"]) for frame in scene.stream(fps=30): frame.tracks # 42 · world-space, id-stable frame.mesh # watertight, 1.9 cm mean error frame.depth("cam-04") # reprojected per-view
First frame
1.4 s
Steady state
11 ms
Calibration targets
0
What a frame carries
frame.tracks 42 world-space, ID-stable
frame.mesh watertight, 1.2 M verts
frame.depth(cam) per-view, 16-bit
frame.occupancy 10 cm voxel grid
frame.t fused clock, ±2 ms
Benchmarks
Numbers first — the harness ships with the docs.
Measured on the junction reference set — 8 cameras, 4K, mixed daylight and sodium-vapour night. Single node, batch 1, no frame dropping.
End-to-end latency
11 ms
p50, steady state
Mean geometric error
1.9 cm
vs surveyed ground truth
Track continuity
99.2%
across full occlusion
Cameras per node
24
2U, 400 W typical
Method
Latency
Mean error
Occlusion recovery
Live
HyperFUSE
11 ms
1.9 cm
99.2%
Live
Offline photogrammetry
4.2 min
0.8 cm
99.8%
Batch
Single-view depth
8 ms
31 cm
42.0%
Live
Multi-view stereo, per-pair
140 ms
6.4 cm
71.5%
Live
Occlusion recovery — share of tracks held through a full-body occlusion longer than 1.5 s.
One scene, every target
The scene is the product. Render it anywhere.
veh-04
ped-11
ped-19
veh-07
cyc-02
ped-28
veh-13
3D view
scene-graph · world frame · 30 fps
3D view
Watertight mesh plus world-space tracks. Query by volume, not by camera — "who is inside bay 4" is one call.
Vertices 1.2 M
Tracks 42
Mean error 1.9 cm
Runs on
UnityUnreal EngineWebXR
Emit
frame.mesh.to_glb()
Temporal fusion
Geometry accumulates as the clip plays.
Every frame refines the model instead of replacing it — occluded surfaces stay solved once any camera has seen them.
t
4.1 s
Vertices solved
434,400
Coverage
45%
Deployment
Runs where the cameras are.
Frames are the sensitive asset — most sites cannot ship them off-premises, and the ones that can still cannot afford the round trip. HyperFUSE runs as an on-prem node, in a regional cloud, or split across both with automatic failover.
Fusion node beside the NVR On-prem
Frames never leave the site. 24 cameras per 2U node, sized in the LENS View calculator.
Cloud in the contracting region Regional
Pune for India and ASEAN, with US/EU and GCC regions on the same image.
Local fusion, central query Hybrid
Geometry stays on site — only the scene graph replicates. Cuts egress to under 1% of raw video.
Reference node
One appliance beside the NVR. Frames never leave the room they were captured in.
Security posture
Frame retention Off by default
Data residency Per entity
Encryption at rest AES-256
RBAC + audit log Included
SOC 2 Type II In progress
In production
Four deployments already running on LENS infrastructure.
Junction reconstruction
Vehicle and pedestrian paths in world coordinates — near-miss geometry, not just counts.
Junctions live 12
Near-miss recall 94.1%
Site-wide volumetrics
Stockpile volume, exclusion-zone breach, and crane sweep from the cameras already installed.
Area covered 3.1 km²
New sensors added 0
Density and flow
Per-square-metre density from overlapping views — holds through occlusion at festival scale.
Peak tracked 11.4 k
Density refresh 2 Hz
Made in LENS Studio
A flat mark, solved into a real object.
The LENS chevron-E, built from its own spec sheet — four bands, one 29.6° cut, one continuous ramp. Drag to inspect; it downloads as OBJ or GLB with every band named.
Drag to orbit · scroll to zoom · right-drag to pan
Chevron-E — from the brand spec
Bands 4
Cut angle 29.6°
Glyph width 0.769 × cap
Warm ramp FF2B44 → FF6A25
Deep teal #005057
Exports OBJ · GLB
Why the mark
The E is shared with LENS — the same glyph on every LENS letterhead. Studio rebuilt it from ratios, not pixels, so it holds at any scale in any renderer.
Open the SDK docsLENS Studio New
3D scenes as code — inside HyperFUSE XR.
A 2D frame goes in. An animated, embeddable 3D scene comes out — motion, sound, and logic all declared in the same file.
2D image → 3D object One frame in, a solved mesh out — the same reconstruction core HyperFUSE runs live.
Timeline animation Keyframes and easing declared in the scene file, not clicked into a timeline UI.
Sound & reactivity Bind any material or transform to an audio band and it moves with the track.
Physics & motion Rigid bodies, constraints, and collisions solved against the reconstructed geometry.
Logic & node scripting Graph or code — both compile to the same scene, both run in the browser.
Generate from a prompt Describe the object and Studio drafts the parametric scene for you to edit.
Export GLB / USDZ / video Named parts and materials survive the export — Blender, Unity, and Unreal open it clean.
Real-time collaboration Scene files are text, so review is a diff and edits merge like any other code.
Publish to web & AR One embed tag renders in the page, on a headset, or in the phone camera.
Get an API key.
Free through 50,000 frames — keys issue instantly. Bring three camera URLs and a terminal.