It is 2009, and photo-sharing sites hold something no survey ever produced: a hundred and fifty thousand tourist photographs tagged 'Rome' — every angle, every season, every camera, no calibration, no order. Reconstruct the city in three dimensions from that pile in a day: match features across wildly different views, chain matches into relative camera geometries, and jointly optimize millions of 3D points and tens of thousands of camera poses in one vast nonlinear least-squares problem — engineered to run distributed, seeded with outlier matches that must not poison the solution. Get it wrong and the Colosseum folds into itself from one mismatched arch repeated eighty times; get it right and 3D reconstruction stops requiring surveyors — the world's casual photographs become its map.
LeCun's convolutional networks, developed from the late 1980s onward, are built for learned feature extraction and recognition from images, which touches this problem's feature-matching sub-task only distantly — the actual 2009 pipeline used hand-engineered SIFT descriptors, not learned CNN features, and LeCun's own research centers on recognition rather than multi-view geometric reconstruction or bundle adjustment. His energy-based models and later self-supervised work show broader interest in structured prediction, but none directly address structure from motion, leaving the connection at the level of shared territory rather than applicable method. His broader vision-research career and his own graduate work on efficient, structured computation for perception give him general sympathy for the correspondence-and-geometry problem here, even though the actual 2009 pipeline was built by researchers working in a hand-engineered-feature paradigm his later career moved away from.
Allen's pioneering work on optimizing compilers at IBM, the first woman to win the Turing Award, gives her deep expertise in restructuring large computations for efficient execution — a skill this problem needs to make a bundle-adjustment computation with millions of unknowns run distributed rather than sequentially. Her program-analysis techniques for parallelizing and optimizing large codebases are conceptually adjacent to the systems-engineering challenge of scaling structure from motion across a cluster. She has no background in computer vision, geometric estimation, or the specific mathematics of camera calibration and feature matching, so the reconstruction's actual content is foreign to her, and her contributions would be confined to the computational-infrastructure layer rather than the geometry the problem centers on.
Battle #80 · 8/10/2026, 11:36:33 AM · this result is deterministic: the same two personas on this problem always resolve the same way.