🚧 This website is under construction and will change shortly. Thank you for your patience!
🎓 Student Project · UC Berkeley · Summer 2026 — an independent student project, not an official UC Berkeley website and not a commercial service.
BB
Berkeley Box Builders University of California, Berkeley
Demo

Photogrammetry · Computer Vision · UC Berkeley

From a simple video to a 3D model
with true dimensions.

Film an object with your phone, upload the footage, and receive a measurable 3D model — calibrated to real-world scale with millimeter-level references. Measure width, height and depth as if you were holding a tape measure.

See how it works

No app install · runs in your browser · processing on our reconstruction server

📐

Real-world accuracy

A printed calibration board anchors your model to metric scale — dimensional accuracy of typically 1–3 % with a careful capture.

📱

Any smartphone

No LiDAR, no special hardware. One slow walk around the object with your phone camera — or a set of 30+ photos — is all it takes.

📦

Box dimensions instantly

Width, height and depth of the smallest enclosing box, in metric or US units — plus free point-to-point measurements directly on the model.

⬇️

Standard formats

Download your model as GLB (glTF binary) in real-world meters, ready for CAD, Blender, AR viewers and web embedding.

How it works

Four guided steps — around 15–30 minutes from upload to measurable model.

  1. 1

    Capture

    Set the object down so it stays perfectly still, place our printed calibration board next to it, and walk slowly around it filming for 30–90 seconds — the camera moves, never the object.

  2. 2

    Upload

    Drop your video or photos into the scanner. Choose your scale reference — the checkerboard is detected fully automatically.

  3. 3

    Reconstruct

    Our pipeline recovers the camera path, computes depth for every view and fuses everything into a single 3D surface — on GPU.

  4. 4

    Measure

    Inspect the model in 3D, toggle the bounding box for width × height × depth, measure any two points, switch between cm and inches, export GLB.

Built on proven computer vision

An open, inspectable reconstruction stack — the same class of technology used in surveying, VFX and cultural-heritage digitization.

Structure-from-Motion

COLMAP 4.1 recovers every camera position and a sparse 3D skeleton of the scene from image features alone.

Dense stereo on GPU

PatchMatch multi-view stereo computes per-pixel depth and fuses millions of points into a coherent cloud (CUDA).

Surface reconstruction

Open3D Poisson meshing turns the point cloud into a clean, watertight surface with vertex colors.

Metric calibration

The printed checkerboard is triangulated across dozens of views; a similarity transform locks the model to real-world millimeters.

Honest about accuracy

Photogrammetry from a plain video is scale-ambiguous by physics — that is why a printed reference is part of the workflow.

What you can expect

  • Typically 1–3 % dimensional accuracy with the calibration board
  • Best results on matte, textured objects in even light
  • Full 3D surface with photographic vertex colors

Known limitations

  • Shiny, transparent or textureless surfaces reconstruct poorly
  • Not a substitute for certified engineering metrology
  • Processing takes minutes, not seconds — real geometry is computed

Ready to try it with your own object?