I have a philosophical question of sorts for users familiar with a multi-view stereo setup. Imagine that I need to measure the dimensions of a large building and I take several photos of the building from different viewpoints. After calibrating the camera (intrinsic and extrinsic), I compute matches between points in these images and map them to 3D points. The output of this entire process is a cloud of 3D points. If I specify a scale, origin and co-ordinate axes, then I can perform measurements between points in this cloud which relate to the real-world scene (the building) that I captured. My question is this: how do I determine how accurate was my reconstruction.

Basically, this camera apparatus forms a measurement device. To compare a measurement, one typically needs to use another measurement device that is much more accurate that it. How do I measure the accuracy of a such a system?


From a more pragmatic than philosophical perspective, possibilities include:

Comparing with ground truth data--these data could come from surveying's instruments, or be assumed to be "reliable" (e.g. you could assess the system's accuracy on the Eiffel tower and other well-known buildings for which measurements have been curated);

Verifying physical properties are met (e.g., floors are horizontal, walls are perpendicular, distances between floors or between windows are constant);

Leaving one ore more views out of the 3D estimation process and assessing how far the projections of the estimated 3D points are from the actual 2D positions in these views (ie., jackknifing).

  • $\begingroup$ I see.. I have been looking at ground-truth data but most of the datasets are done using a laser scanner which themselves have an accuracy of the order of a well calibrated multi-view stereo system (according to literature). $\endgroup$ – Mustafa Jan 27 '12 at 21:05

To start with, you might try a reconstruction of different object coordinates from each pair (or small subset) of photos. Then, given a large enough set of reconstructions using different pairs or subsets, look at the statistical distribution and variance of all the different reconstructed coordinates.


Your Answer

By clicking “Post Your Answer”, you agree to our terms of service, privacy policy and cookie policy

Not the answer you're looking for? Browse other questions tagged or ask your own question.