Photogrammetry & Its Limits
About this app
Many photos become 3D: a ring of cameras triangulates the same feature point from different viewpoints and a point cloud grows. Switch matte / reflective / transparent — on shiny and glassy surfaces the reconstruction visibly falls apart. The motivation for neural rendering.
How photogrammetry works
Photogrammetry creates 3D models from many photos of an object. It does not measure shape directly; it finds the same feature point in images taken from different angles and triangulates its position from the intersecting view rays. Repeating this for many points produces a point cloud. This only works if features look the same from every direction, which is the case for matte, textured surfaces. Reflective surfaces show a different image from each angle, and transparent ones offer almost no stable features. The rays then contradict each other and the reconstruction gets holes.
What the app shows
A ring of cameras surrounds an object. The pipeline runs in three steps: capture, matching and reconstruction. View rays show how several cameras aim at the same feature point, and the point cloud grows. The surface can be switched between matte, reflective and transparent. A coverage indicator rates the result, for example “dense and clean” or “mostly failed”.
Try it yourself
Run the reconstruction first with a matte surface and watch the rays meet in one point. Then switch to reflective: each camera aims at a different apparent point and the rays diverge. With transparent, hardly any reliable points remain. This limit is one motivation for neural rendering methods such as NeRF.
Frequently asked questions
- What is photogrammetry?
- Photogrammetry computes the three-dimensional shape of an object from several photos. Matching feature points are found in different images and located in space by triangulation. The result is a point cloud or a mesh built from it.
- Why does photogrammetry struggle with shiny surfaces?
- Reflective surfaces show different reflections from each viewing angle. A point therefore looks different in every photo, and matching between images fails or goes wrong. Transparent surfaces have hardly any features of their own. Both cases lead to gaps and errors.
- How many photos does photogrammetry need?
- Each point must be visible in at least two images; in practice many more are used, with large overlap between neighbouring shots. The exact number depends on object size, shape and the accuracy required.
- Which surfaces can I compare in the app?
- Matte, reflective and transparent. The same pipeline runs for each, and the coverage indicator shows how complete the point cloud becomes.
Subject: Computer Vision
More from Computer Vision
- 3D Gaussian Splatting
- Convolution playground
- CV Foundational Skills
- Industrial lighting
- Neural Rendering (NeRF)
- Pixel grid
- Stereo & ToF — depth vision
A VisuApp by heyprof: interactive, free in your browser, no sign-up. What is a VisuApp? · All apps