tart All work

PERSONAL / INTERACTIVE 3D

Shadow Room

Your webcam silhouette casts a real shadow into a dark 3D living room. The view follows your head, and a cat chases the shadow of your hand.

Personal side project · Interactive 3D & computer vision

My contribution

Built the three.js room and its lighting, a webcam-to-shadow pipeline with MediaPipe in a Web Worker, head-coupled perspective with look-around and calibration, a path-finding cat, and a tracking view, tuned for low lag and low GPU power.

Download CV (English)

Try it · mini version

See what it can do

Shadow Room living room lit from the hallway, with a person's shadow on the curtains and a ginger cat asleep on the sofa
The room at rest: a light in the hallway behind you throws your shadow onto the curtains. Captured from the running app with a simulated person in place of a webcam.

Personal side project · Interactive 3D, webcam tracking & live wallpaper

What this project does

Shadow Room is a live 3D wallpaper of a dark living room. MediaPipe finds your silhouette in the webcam image, and a light behind you casts it as a real, soft-edged shadow across the curtains, sofa and floor. Objects react when your shadow touches them, and small head turns look all around the room.

Inside the project

  • Touch things with your shadow: swat the ceiling lamp and every shadow swings, hold your head's shadow on the TV or the light switch, and ripple the curtains.
  • Look around with small head turns: 10° looks to the side and 15° straight behind you, so you can keep watching the monitor. A Center button sets your normal sitting pose.
  • A procedural cat sleeps on the sofa, chases the shadow of your raised hand, plans paths around the furniture with A*, and bolts when your shadow looms over it.
  • Pose detection runs in a Web Worker; while nothing moves, the room drops to 15 fps and reuses its shadow maps. On the test laptop, that cut the power the page adds from about 11 W to about 4 W while you sit still.
  • A tracking view shows the analysed frame, the mask and skeleton, and per-stage timings, so lag can be traced to the camera, the tracker or the GPU.

Skills & practical experience

Practice turning noisy pose estimates into stable, low-latency interaction, keeping a real-time WebGL scene cheap enough to run all day, and building diagnostics that make performance problems visible.

Three.jsWebGLMediaPipeComputer visionJavaScriptWeb WorkersPerformance tuning

The images and the tour were recorded from the running app in headless Chrome, using its built-in demo figure or a simulated person fed through the tracker; none is webcam footage. The power figures were measured on one laptop with an RTX 3060 Laptop GPU.

Project context

Everything runs in the visitor's browser, and webcam frames never leave the page. It works as a web page or as a Lively Wallpaper desktop background. Without a webcam, a built-in demo figure casts the shadow.

Project media

A ginger cat chasing the shadow of a raised hand across the living-room floor
The cat chases the shadow of a raised hand like a laser dot, then crouches and pounces. Demo figure.
View down the hallway behind the viewer, toward the lamp that casts the shadow
A 15° head turn looks straight behind you, down the hallway to the lamp that casts your shadow. Simulated head pose.
The living room lit by the TV after the ceiling lamp was switched off
Hold your head's shadow on the TV or the light switch: here the TV is on and the ceiling lamp is off. Demo figure.
Tracking view panel with the analysed camera frame, a top-down map and per-stage timings
The tracking view (V) shows each stage of the pipeline. Simulated person; the render and camera rows are live measurements.
34-second tour recorded from the app: the cat's chase, looking around with small head turns, the lamp swat, and the TV. Demo figure and simulated head poses; no webcam footage.