Nils Schnorr and Thomas Leimkühler trained a neural network to separate stones in Carian fortifications. Their October study combines photographs with maps of which way the wall’s surface faces.[1]
Across 663 annotated stones on four held-out walls, the appearance-only model merged 42% with neighbouring blocks; adding geometry reduced that to 26%. The combined model individually resolved 48%, under a test requiring both correct coverage and separation to reach 90%. Finely jointed walls remained difficult.[1]
Labelling pixels and identifying objects answer different questions. Computer vision calls the first task semantic segmentation: which pixels belong to a class? Instance segmentation asks which belong to each separate object. Saurabh Gupta and colleagues made this distinction explicit in their 2014 work on colour and depth images. Recognising a field of bottle pixels, for example, does not tell a robot where one bottle ends and the next begins.[2]
Geometry provides another clue. A change in brightness might be paint or shadow; a change in surface orientation describes shape. The earlier work combined colour with depth and orientation cues to detect contours, including edges that appearance-based methods missed. Geometry supplies evidence about the surfaces themselves.[2]
The gain lies in boundaries, even when most pixels already carry the right label.[1]
Sources & further reading
- Appearance Meets Geometry Journal of Computer Applications in Archaeology · 14 Oct 2026 · Accessed 29 Oct 2026
- Learning Rich Features from RGB-D Images for Object Detection and Segmentation ECCV 2014 · Accessed 29 Oct 2026