IIT Techfest 2026 · NammaSpace challenge

Walk through a real room, rebuilt from one phone video.

NammaSpace turns an ordinary phone walkthrough into a photo-real 3D copy of the space that runs in your browser, with collision so walls and furniture stop you. No install, no scanner, no hand modelling.

Walk through the room → See how it works
27.14 dBheld-out PSNR, room scene
1 videoper scene, nothing merged
10 cmcollision grid resolution
A bedroom rendered from the 3D Gaussian splat in the browser viewer
Live render in the browser, not a photo
What this achieves

A walkable digital twin, not a 3D model

A normal 3D model is built by an artist or a mesh generator and is lit by guesswork. NammaSpace fits half a million translucent 3D Gaussians directly to the real video frames, so the room keeps its real colour, reflections and soft edges.

01

Learned from photos

Every Gaussian is optimised until its renders match the original frames. We hold frames back and score the render against them, so quality is measured rather than claimed.

02

Looks and physics are separate

The splat draws the room. A separate 10 cm collision grid, built from the splat and the camera path, decides where you can stand. You cannot walk through the bed or the wardrobe.

03

Runs where judges are

Each scene is a single 8 to 11 MB file streamed to a plain web page, drawn by our own WebGL2 splat renderer. Open a link and walk.

An ordinary 3D model

  • Surfaces and textures made by hand or by a mesh generator
  • Lighting is approximated, reflections are baked or missing
  • Looks like a model, however well it is made

NammaSpace

  • Appearance learned from the real frames, view by view
  • Soft edges, glossy floor and fabric detail come from the data
  • Collision comes from the scan itself, not a hand-placed box
How it works

From a phone video to a place you can walk in

Four stages, one command each. Plain description on top, what actually runs underneath.

1. Phone video

Walk slowly around the room while filming.

Sharpest-frame picker keeps the crispest frame in each half-second window. Blurry frames are dropped.

2. COLMAP

Work out where the camera was for every frame.

Structure-from-Motion solves each camera pose and a sparse point cloud. Poses are final and never re-run.

3. 3D Gaussian Splatting

Grow a cloud of soft 3D blobs that reproduce the real views.

nerfstudio splatfacto, 30,000 steps, full-size frames, trained on an RTX PRO 6000 Blackwell.

4. Walkable digital twin

Add floor, walls and furniture collision, then publish.

Floor-aligned and scaled, 10 cm grid with a 25 cm body margin, W A S D controls in a WebGL2 viewer.

Evidence

Three experiments, measured

Quality is scored on frames the model never saw during training. Figures come from each scene's training report.

SceneFramesGaussiansPSNRSSIMLPIPSTrainingWalkable floor
Early bedroomnot recorded700,000 shown––––8.1 m²
Hallway and bedroom868650,65426.46 dB0.8740.244995 s15.8 m²
Room613541,33727.14 dB0.9130.191819 s4.3 m²

PSNR and SSIM: higher is better. LPIPS: lower is better. Scores are measured on frames held out from training. Walkable floor is the area your body centre can reach after the 25 cm margin, so it is smaller than the room. The early bedroom predates our evaluation step, so no held-out figures exist for it.

Hallway capture: held-out PSNR as the pipeline improved

Hallway v1, 709 frames23.56
Hallway v2, 868 frames26.46
Room v1, sharpest frames27.14

Bars scale from 0 to 30 dB. Room v1 is a different scene, shown for pipeline progress rather than a like-for-like comparison.

The viewer

Built to be explored, not just watched

The viewer with the live map at the bottom left, controls strip and display modes
The viewer in the room scene
Close view of the live map showing walkable floor in teal, obstacles in amber, and the player position with viewing cone
Live map and controls, close up
  • Live map on the leftDrawn from the same collision grid the walls use. Your position, viewing cone, walkable floor and obstacles update as you move. Press N to hide it.
  • First-visit guideA short card explains movement, look, run, jump and the view modes. Reopen it from the Guide button.
  • Inspect the collisionPress M to overlay the collision grid in the 3D view. Press 2 to see the scan mesh where a scene has one.
  • ControlsWASD walk, Shift run, Space jump, mouse to look, R back to the door.
The scenes

A progression of experiments

Each scene is one video and one reconstruction. We kept the earlier ones online so you can see what changed.

Early bedroom reconstructionExperiment 1 · early

Early Bedroom

Our first bedroom reconstruction. The ceiling and distant walls are rough, which is the baseline we improved on.

Splats
700,000
Collision
52 × 71 cells
Open →
Hallway and bedroom reconstructionExperiment 2 · longer capture

Hallway and Bedroom

A longer walk from a hallway into a bedroom. The hallway is excellent, the bedroom end is rougher.

Held-out PSNR
26.46 dB
Collision
89 × 74 cells
Open →
Room reconstructionExperiment 3 · best

Room

A real bedroom from one phone video with the sharpest-frame picker. Our best quality scene.

Held-out PSNR
27.14 dB
Collision
58 × 53 cells
Open →
Honest limits and next steps

What works today, and what comes next

Known limits

  • Ceilings and areas the camera barely saw are rough or dark.
  • Collision is a 2D floor grid. It stops you at walls and furniture but does not model stairs or overhangs.
  • Scale assumes a fixed phone height rather than a measured reference object.

Next

  • Points of interest placed on the same grid the live map already reads.
  • A* route finding between points of interest, drawn on the live map and in 3D.
  • Stronger collision where one source of evidence disagrees with the other.