Outline the classic image stitching pipeline.

WHAT THIS TESTS The interviewer wants the canonical multi-stage geometric vision pipeline and the right tool for each stage, plus awareness that real matches are noisy.

A GOOD ANSWER COVERS Stage one is feature detection and description: find distinctive, repeatable keypoints with a detector like SIFT, SURF, or ORB and compute invariant descriptors so the same point can be recognized across views despite scale and rotation changes. Stage two is matching: pair descriptors between the two images, typically by nearest neighbor with Lowe's ratio test to discard ambiguous matches. Stage three is robust transformation estimation: because the views are related by a projective homography for a rotating camera or planar scene, fit a homography from matched points, using RANSAC to tolerate the many wrong matches that survive. Stage four is warping: project one image into the coordinate frame of the other using the homography. Stage five is compositing: correct exposure differences and blend the overlap with feathering or multiband blending to hide the seam.

COMMON WRONG ANSWERS Assuming a simple translation or affine model when a homography is needed. Omitting RANSAC, so a few bad matches corrupt the fit. Forgetting blending, leaving a visible seam or exposure step.

LIKELY FOLLOW-UPS When a homography is valid versus when you need full structure from motion, how RANSAC chooses inliers, cylindrical or spherical warping for wide panoramas, and bundle adjustment for many images.

ONE CONCRETE EXAMPLE For two photos of a building taken by pivoting the camera, SIFT finds corner-like keypoints on windows, ratio-tested matches link them, RANSAC fits a homography from four-point samples while rejecting mismatches, the second image is warped onto the first, and multiband blending merges the overlap into one seamless panorama.