Guiding notes for photographing for 3D reconstruction

Written by Paul Bourke
August 2013

The following gives some guiding principles on photography for 3D reconstruction purposes, that is, automatically reconstructing a textured 3D model from photographs. The software upon which these principles have been tested are: Bundler based, 123D Catch, and PhotoScan. There is no substitute for experimenting and gaining a feel for the best approach, but hopefully the following will provide a short-cut.

  • Use a fixed focal lens, otherwise known as a "prime lens" for an SLR camera. Most point-and-shoot cameras have a fixed focal lens. Note, this is generally not advice, but a requirement.

  • Never take two photographs from the same location.

  • One will never reconstruct parts of the object that are not captured in a photograph.

  • Generally move around the object in approximately 10 degree steps, example 1 below. Repeat at one or more levels if the object is concave vertically, example 2 below.

  • 2.5D objects, for example engravings and height fields, are significantly easier and usually require fewer images, example 3 below.

  • The examples shown here are for self contained objects, generally try to keep the object filling each frame.

  • The guidelines become more difficult and more software dependent for linearly scanned objects where the object in question does not fill the photograph frame. At least 50% overlap is suggested, more if the texture is subtle. See example 4.

Example 1: Rocks outside Geology building, UWA


Cameras (blue rectangles) circle the central rock in approximately 20 degree steps.

Example 2: Lion statue, Hong Kong Dragon Gardens


More complicated object requires photographs (red) at different heights to capture internal convolutions.

Example 3: Temple façade, Gommateswara, India


2.5D panels require the fewest photographs. Shown here in a circular arc but a linear sweep would be just as good.

Example 4: Dragon staircase, Hong Kong


2.5D example but illustrating overlapping linear photographic track. Each shot only has a portion of the object in view.