Human presence, translated into a living 3D forest.
Data Forest is a real-time installation that turns building occupancy into a visual environment. Camera data feeds directly into tree growth, particle density, wind, weather, and atmosphere inside a Unity-rendered forest.
More people, more life. A quiet building makes a sparse, still scene. A busy one fills the canopy, adds wind, and triggers atmospheric events. The forest mirrors the building in real time.
There is a particular feeling I have been trying to reproduce. If you have ever stood inside a building, a campus, or a compound and come across a scale model of that same place, you know it. Mine came at the exit of an old lead mine: after an hour underground we surfaced next to a 3D map of the entire mine, and the tunnel we had just walked through was reduced to a few inches of bent wire. For a moment you hold both scales at once. You feel small, and you feel connected to the whole structure, because seconds ago you were inside the object you are now looking down at.
Data Forest starts from that feeling and adds the thing a static model cannot do: it updates itself. The forest is a live portrait of the building, drawn continuously by the people moving through it, and nearly all of them are unaware of it. Someone crossing a hallway two floors up adds growth to their zone of the forest and stirs the canopy without ever knowing the piece exists. When the building fills, the forest thickens and moves. When it empties, it goes still.
The viewer completes the loop. A camera near the screen reads the people standing in front of it, so a visitor is simultaneously the audience for the portrait and one of its subjects. You are watching a crowd you belong to, and your presence is changing the image while you watch it. That is the strange doubling the mine map only hinted at.
The last intention is the record. Crowds are rare, short-lived events: a building fills, hums for a few hours, and empties without leaving a trace. The forest gives that energy somewhere to accumulate. A full house shows up as growth and motion, and the quiet afterward reads as stillness. In that sense the piece is less a data display than a way of celebrating, and briefly remembering, people gathering.
Person counts get smoothed into a biomass value for each micro-zone. That value drives the growth stage of the custom SpeedTree oak, grass coverage, and how fast things grow or decay. Each zone responds to its own camera independently.
Noise values pipe into the particle system and drive the firefly-like elements, drifting motes, and ambient glow. Even when the building is quiet, the forest stays alive.
Movement metrics connect to wind intensity. Trees and grass share the same wind direction and rate, so activity in the building translates into one coherent physical response across the scene.
There's also a more direct layer. A camera near the installation reads viewer poses and triggers live responses in the render, including rain and audio. The rain works with a reactive sky asset, so people standing in front of the piece can actually change the weather in the scene.
Webcams observe selected zones and stream frame data to the detection server.
Python, OpenCV, and YOLOv8 process frames and compute per-camera people counts.
FastAPI exposes per-camera values, aggregate counts, smoothed readings, and system health.
Unity polls the HTTP endpoint and converts current metrics into normalized control values.
Micro-zone controllers drive tree growth, grass, particles, wind, weather, sky state, camera behavior, and ambience.
Captures webcam frames, detects people, and outputs clean metrics without storing any identity data.
A lightweight HTTP server that exposes the current system state. Easy to inspect in a browser and simple for Unity to poll.
Unity owns the live 3D world. C# scripts take incoming metrics and drive biomass, particles, wind, rain, camera movement, and audio.
A custom SpeedTree oak, grass, rain systems, particle effects, reactive sky, and scene lighting. The goal was for it to feel like a terrarium you're looking into, not a data readout.