Multi-View Stereo Quiz

Answers and explanations appear as you take the quiz.

1. What does Multi-View Stereo typically assume is already known before it starts?
   • The camera poses for each image
   • The final 3D mesh
   • The colors of every surface
   • The names of objects in the scene
2. What core principle does MVS use to find correct 3D points?
   • Maximizing image brightness
   • Photo-consistency across multiple views
   • Minimizing file size
   • Random sampling of pixels
3. How does MVS differ from the output of Structure from Motion?
   • MVS only works on a single image
   • MVS ignores camera positions entirely
   • MVS produces dense per-pixel depth, while SfM gives a sparse point cloud
   • MVS outputs audio instead of geometry
4. What does plane-sweep stereo do?
   • Rotates the camera physically
   • Sweeps dust off the lens
   • Sorts images alphabetically
   • Sweeps a virtual plane through depth and finds where views agree best
5. Which surfaces are especially challenging for MVS?
   • Textureless and reflective surfaces
   • Brightly colored painted surfaces
   • Rough stone surfaces
   • Anything with sharp edges
6. How do learning-based methods like MVSNet improve on classic MVS?
   • They remove the need for any cameras
   • They learn matching costs and depth regularization end-to-end for greater robustness
   • They only work in grayscale
   • They produce 2D images instead of 3D
7. How may an MVS system compare a hypothesized depth across neighboring views?
   • Compare only the file names
   • Use the language-model perplexity
   • Warp image patches and measure their appearance agreement, for example with normalized cross-correlation
   • Count camera metadata fields
8. After per-image depth maps are estimated, what can MVS do with them?
   • Convert them into language tokens
   • Replace the camera poses with labels
   • Discard all cross-view information
   • Fuse them into a denser point cloud or mesh and filter inconsistencies