How photogrammetry works
Photogrammetry reconstructs an object from many photographs taken from different positions. The software first finds distinctive features in every image, then matches the same features across images. From those matches it works out where each camera stood, a step called structure from motion, and produces a sparse cloud of points.
It then estimates depth for each image, fuses the depth into a dense mesh, filters it, and projects the photographs back onto the mesh as its texture. Meshroom’s default pipeline lists exactly these stages, from feature extraction and matching through structure from motion, depth maps and meshing to texturing.
Objects that scan well, and objects that do not
The object has to stay the same shape for the whole shoot, so rigid objects work and soft, bendable or articulated ones cause failures. Its surface has to give the software features to match: textured, matte materials reconstruct well.
Apple’s capture guidance advises avoiding objects that are very thin in one dimension, highly reflective, transparent or translucent, and warns that single-colour or very smooth surfaces may not give enough data. Glossy plastics, chrome, glass and clear packaging, which are common on consumer products, are exactly the materials photogrammetry handles worst.
- Scans well: rigid, matte and textured surfaces, such as stone, wood, unglazed ceramics, fabric-covered hard goods and many foods.
- Scans poorly: shiny, transparent, translucent, very thin, or a single plain colour.
- Fails: anything that moves, bends or changes shape between shots.
Taking the photographs
Either walk the camera around the object at several heights, or put the object on a turntable in front of a plain, well-lit background and rotate it. Turn the object over between passes so the base is captured too. Apple recommends that neighbouring shots overlap by 70% or more, and warns that under 50% may make reconstruction fail.
Fill the frame with the object without cutting any of it off, use a narrow aperture for depth of field, and shoot at full resolution, in RAW if possible, from a tripod. Keep focal length, aperture, shutter speed and ISO the same for every shot. Light it with soft, diffuse light: hard light from a flash or direct sun casts shadows that confuse the reconstruction and end up in the texture. The Meshroom tutorial calls the shoot the most important and challenging part of the whole process.
From raw scan to usable model
A raw reconstruction is dense and uneven, may have holes where the cameras could not see, and carries the shoot’s lighting in its colour texture. Scale is another gap: if the photos carry no depth data, the model has no real-world size and has to be scaled against a known measurement.
Clean-up usually means filling holes, removing stray geometry and setting the scale, then reducing the mesh. Remeshing suits a static object; a model that has to animate or run in a viewer is normally retopologised and its detail baked into normal maps. The texture may then be rebuilt so lighting can be added in the render instead of being fixed in the photographs.
Photogrammetry for product CGI
Photogrammetry suits products that exist only as physical objects, such as heritage designs, handmade pieces or organic forms that would be slow to model. It records an object as it is, including wear and imperfections.
For a product launch that is often the wrong starting point. Most products already exist in CAD, which carries the exact dimensions and the finish as designed, and many are glossy or transparent. In those cases a model built from the CAD, with photographs used as reference for colour and materials, is the more reliable route.