Skip to main content

Video-to-3D Reconstruction

Upload a video and reconstruct a 3D scene from it using a multi-agent ML pipeline.

Status: planned. This feature is not currently in the codebase (prototype removed 2026-07).

The sections below describe the intended design. No video store, panel, or startVideoReconstruction Cloud Function currently ships.

Pipeline Stages

StageML ModelPurpose
segmentSAM2Segment video frames into objects
depthMonocular depth estimationEstimate per-pixel depth
classifyObject classificationIdentify what each segment is
rigAuto-riggingAdd armatures to characters/vehicles
composeScene composerPlace objects in 3D space

Job Lifecycle

uploaded → segmenting → depth → classifying → rigging → composing → done
└→ failed

Each stage would update a videoJobs/{jobId} Firestore document, and a VideoPanel would subscribe via onSnapshot to show real-time progress bars.

UI

A VideoPanel would provide:

  • File input for video upload
  • Job list with per-stage progress indicators
  • "Reconstruct" button to kick off the pipeline
  • Result GLB download when complete