DHUM250@UVic

Things to Think With: Introduction to Modelling and Printing 3D Objects (Spring 2014)

View project onGitHub

Photogrammetry

Photogrammetry is the process of making measurements and stitching together models using photographs.

In order to build a model using images, you'll need a camera and a "reference object" you want to digitize and model. Any digital camera will work, really. It does not need to be, say, a fancy DSLR. You'll then use that camera to take multiple photographs of a single object. Those photographs will then be fed into a computer vision algorithm (e.g., 123D Catch or VisualSFM), which will stitch the images together for you.

As you begin to photograph your reference object, here are a few tips:

  • KNOW YOUR ALGORITHM. Before you spend too much time on a particular tool, kit, or app, see how others have used it and what guidelines exist. For my work, I use either 123D Catch or VisualSFM. 123D Catch is very user-friendly, but limits your choices and renders many of its own choices opaque. VisualSFM has a slight learning curve, runs in part via the command line, and affords more range / choice.
  • Avoid reflective or transparent objects, or lightly dust them with powder.
  • Also avoid reflective walls and floors.
  • Avoid featureless objects.
  • Avoid things that move as well as moving things (e.g., don't move your object as you're photographing it).
  • Avoid direct light. Ambient light is best, and keep the lighting uniform (where possible).
  • Place your object on a circular stage (or the like) that you can move around as you're photographing it.
  • Strategically arrange self-occluding objects so that you can photograph as much of them as possible.
  • Produce reference points (using a newspaper or post-it notes).
  • Do not use a flash.
  • For my camera (Canon EOS Rebel T3i), I have found that an 18mm focal length is best.
  • Whatever the focal length, keep it consistent across the photographs.
  • Fill your photographs with the reference object. Avoid photographing content you don't want in the model.
  • Aim for sharp images wherever possible.
  • Move steadily and consistently around your object.
  • When photographing, use tight intervals for occlusions.
  • Try taking two "loops" around your object, and photograph from at least two different angles.
  • Take your detail shots last.
  • For most computer vision algorithms, between 40 and 60 photographs is best, unless your object has a number of occlusions. For small objects, you might only need between 20 and 30 photographs.

When working with 123D Catch, I've found this video informative:

You might also want to check out the following:

Curious (more generally) about computer vision as a cultural formation? Then you might appreciate this video, by Timo Arnall: