GAD Editorial · Published · · English study · Chinese summary · Study dated 2026-10-06
An adaptive-reuse survey often begins with a practical question: which existing objects deserve closer inspection before the fit-out is removed? Searching a captured interior for chairs, cabinets or washbasins could help organise that work. The important deliverable, however, is a reconciled list of identifiable objects, each linked to evidence and a follow-up decision.
OpenMask3D is research in open-vocabulary 3D instance segmentation. It combines predicted object masks with image features gathered across views, allowing free-form queries to retrieve separate objects. Its demonstrations include object shapes, apparent materials and possible uses. This supports a proposed inventory-review exercise; it does not establish that a retrieved object is reusable, correctly identified or even fully captured.
Begin with one surveyed room and an ordinary manual inventory. Give every visible object a stable reference, then connect that reference to photographs and its approximate position. Record inaccessible cupboards, concealed installations and rooms excluded from capture. This preliminary work makes it possible to distinguish a missing search result from an object that was never represented in the input.
The official repository requires a scene point cloud and posed RGB-D frames for its single-scene pipeline. It also warns that masks produced directly for one scene can differ from those used in the ScanNet evaluation. A published benchmark score therefore cannot stand in for checking the actual capture, its alignment and the resulting object boundaries.
For the proposed pilot, search the same room using several descriptions that matter to the reuse brief. A general query for seating and a narrower description of an upholstered chair should produce reviewable candidate sets. Preserve the exact wording and the returned masks. If two masks describe the same chair, merge the inventory references only after visual confirmation. If a chair disappears under a narrower query, retain the original record rather than treating the new search as a complete replacement.
Reconciliation should work in both directions. Examine each returned object for false matches, and walk through the manual list to find objects the search missed. Pay particular attention to stacked furniture, repeated units and objects partially hidden behind other objects. A tidy candidate gallery can conceal all three. Record corrections per object so that the team can judge whether retrieval actually reduces review effort.
Only then add the information needed for a reuse decision: measured dimensions, ownership, physical condition, disassembly access and any required specialist assessment. A visual match to a material description is not a material test. These fields should identify their human reviewer and evidence separately from the machine-generated suggestion.
The useful handover is a candidate register with confirmed identities, unresolved cases and a fieldwork list. It can inform what the architect asks to preserve and what needs another visit. Success means finding worthwhile inspection targets without losing count of what the scan and search failed to show.