Auto-Stitching Photo Mosaics

Taking two overlapping photos and producing a seamless panorama is a cascade of five hard problems: finding distinctive points (corner detection), distributing them evenly (ANMS), matching them across views (feature descriptors + similarity search), filtering out bad matches (RANSAC), and blending the final warped images without seams (pyramids). Each step uses well-established CV techniques; chaining them is where everything gets interesting.

Feature Detection

Harris corners find distinctive points — but they cluster. Textured regions (like grass or a corner of a building) produce hundreds of nearby strong corners, while other interesting areas get ignored. ANMS distributes corners spatially by keeping only those whose response is stronger than all neighbors within some radius. The result: features spread evenly across the image, which makes downstream matching more reliable.

Feature Matching + RANSAC

Matching features naively gives you hundreds of candidates, most of them wrong. Lowe’s ratio test (accept a match only if its nearest-neighbor distance is much smaller than second-nearest) filters aggressive false positives, but still leaves outliers caused by repeated patterns.

RANSAC is the robust estimator. Sample 4 random matches, compute the homography they imply, count how many other matches agree with that homography. Repeat 1000 times, keep the homography with the most inliers. The green lines in the second panel are the RANSAC-verified matches; the red outliers got discarded because no consistent homography could explain them.

The outcome: a 3×3 homography matrix H that maps points from image A’s coordinate frame into image B’s. Warp image A through H, overlay on image B, blend with a pyramid — panorama done.

Why the Pipeline Works End-to-End

Each stage absorbs noise from the previous. Harris over-detects — ANMS filters. Matching over-connects — ratio test filters. Ratio test still leaves outliers — RANSAC filters. Even if each stage has 50% error, stacking four independent filters reduces the final error rate to ≈6%. The pipeline is deliberately redundant: the whole point of classical computer vision is surviving the failures of individual components.