Clips of 4D reconstruction are everywhere right now: a video turned into 3D frame by frame, each frame left where and when it happened, so moving things leave trails through space. It is a demonstration we built ourselves: no client commissioned it, and anything marked sample is made up and says so on screen.
How it was made
We built our own pipeline from open parts and ran it on a screen recording of our factory showcase. The still hall became a Gaussian splat; every robot and worker that moves was rebuilt in 3D for each of 155 frames. In the browser you can scrub time, run it backwards, freeze a moment and walk round it, or leave ghost trails. A capture report runs underneath: how much of each frame moves, how fast the camera turns, a plan view of where it went, events to jump to, and a build-up of every moving point in the order it was captured.
Match to the video in moving regions: 17.3 dB. PSNR from the original camera, mean of 5 frames
Match over the whole picture: 23.9 dB
Camera solved: 165 of 165 frames. 0.7 px reprojection error
Depth error on held-out points: 2.4%. Median; 3.3% at the 90th percentile
Busiest moment: 1.6 s. 1.6% of the frame moving there; 0.8% on average
Fastest camera turn: 81°/s. At 1.7 s; the camera turns 38° over the clip and never holds still
Processing, 5.5 s clip: about 11 min. Camera solve 5.5 min (CPU), depth 43 s, motion 49 s, splat 2 min 50 s
What it will not tell you
One camera means one side of everything: the viewer keeps you within about 15° of the path the camera flew, because further out the moving things are revealed as painted cut-outs and the hall turns to fog. The source is our own clean 3D render, not phone footage, and processing takes minutes, not real time. A measurement taken from a capture is a guide: where a figure matters legally or commercially, the surveyed number rules, and nothing here replaces an inspection.
4D Video Playback · Spatial Intelligence, Reimagined
Step inside a video, and walk around a moment.
A flat screen recording of our factory cell, rebuilt as a scene you can move through in space and in time. The still world became a Gaussian splat; every person and robot that moves was rebuilt in 3D, frame by frame, exactly where and when it was. Freeze time and walk around it, run it backwards, or leave ghost trails. Underneath, a capture report measures every frame: how much of the picture is moving, how fast the camera turns, where it went, and the moments worth jumping to.
0.00SMode · Now
Frame 1 / –
Drag to orbit · right-drag to pan · scroll to zoom
Loading the 4D scene…
Events
Speed
Time
Colour
Layers
Moving points this frame
–
Recorded so far
–
Coverage so far
–
Moving points, whole clip
–
Downloaded
–
Drawing
–
Capture record
How, where and how well this clip was captured
Source
Screen recording of our factory cellEXP-043, real-time 3D, our own footage
Clip
–
Lens
–
Camera solve
165 of 165 frames placedour camera solve, 0.7 px reprojection error
Depth
2.4% median erroran AI depth model, on solved 3D points held back from the fit
Camera path
–
Busiest moment
–
Camera holds still
–
Moving points
–
Still world
Gaussian splat, 9.0 MBour 3D training, moving pixels ignored
Match to the video
14.8 → 17.3 dBinside moving regions, from the source camera, without and with the rebuilt points
Processing
About 11 minutesone RTX 4070, mostly camera solving on the CPU
How we built it
From a flat video to a scene you can walk through in time
Find the camera
Our camera solve matches features between frames and works out where the camera was for every frame: all 165, with 0.7 pixels of error. Features on moving things disagree with the rest, so the solve rests on the still world.
Estimate depth
An AI depth model made for video gives every frame a depth map that stays steady over time. Its scale is unknown, so each frame is fitted to the camera solve's 3D points, which puts all frames in the same units. On points held back from the fit, the depth is within 2.4% (median).
Find what moves
Each pixel becomes a 3D point, projected into frames up to two-thirds of a second earlier and later. If those frames see straight through the spot where the point was, or see a different colour there, something moved. No segmentation model, so it works on anything that moves, including a robot working in place.
Rebuild the still world
Our 3D training builds a Gaussian splat of the hall from the same frames, told to ignore the moving pixels, so people and robots don't leave smears in it.
Rebuild every moment
The moving pixels of each frame become 3D points in the same space, 10 bytes each: 350,000 points over 155 frames, 3.5 MB. The player keeps them all on the graphics card, so any moment, trail or tube is one draw call away. Seen from the original camera, adding them lifts the match to the video inside the moving regions from 14.8 dB to 17.3 dB.
Report the capture
The player measures the clip as it loads: the share of each frame that moves, how fast the camera turns and travels, and a growing box around everything that has moved so far. Peaks in movement, the fastest turn and any stretch where the camera holds still become events you can jump to, and a plan view shows where the camera was and what it could see. It is a record of how, where and how well the clip was captured, not only the result.
Limits
Everything comes from one camera, so the orbit is kept within about 15° of the path it flew. Further out, the still world turns to fog and the people and robots are revealed as painted cut-outs, because one viewpoint never saw their far side.
The source is a screen recording of our own real-time 3D factory, so it is clean, sharp footage. Phone video adds motion blur and rolling shutter, which make camera solving and depth harder.
Moving things are rebuilt as coloured points each frame, not as tracked, persistent objects. That is what makes trails easy, and also why they shimmer slightly. A few still edges are picked up as moving, and robots that stand still for the whole clip stay part of the still world.
Two of the report's measures are in real units: the share of the frame moving and the camera's turn rate in degrees per second. Camera travel, the plan view and the coverage box are relative, because one camera cannot tell a small room from a large one. "Moving" means the geometry disagreed between frames; it does not say what moved.
Processing the 5.5-second clip took about 11 minutes on one RTX 4070, most of it camera solving on the CPU. This is offline reconstruction, not live.