MessyTable matching benchmark Established November 11, 2026 selection.json defines the preserved scene membership and official source hashes. The same 100 scenes from 100 camera setups are used for all baseline and Step 2 runs. Each scene retains nine camera views and all 36 unordered view pairs. 100 scenes 900 photos 3,600 view pairs 21,706 annotated object appearances 2,484,124 candidate object pairs 78,335 true object pairs 30 easy scenes, 50 medium scenes, 20 hard scenes The source photos remain under data/messytable-full/MessyTable/images/. The manifest is the stable input list. New experiments must verify its SHA256 recorded in their plans and preserve the camera and annotation order. asnet/ and vit/ contain the official-protocol reference runs. step2/ contains the fixed Step 2 variation sweep and its selected top three. Model selection on this benchmark exposes its results to development. This benchmark remains the comparison set for subsequent experiments.