TRACE: Instance-Level Open-Vocabulary Inventory Generation for 3D Forensic Evidence Reconstruction

Abstract. State of the art crime scene documentation not only relies on visually and geometrically faithful 3D reconstruction, but also structured access to potentially relevant evidentiary objects. Existing language-aware 3D scene representations enable open-vocabulary object querying and scene-level reasoning, yet they do not explicitly tackle the formation of globally consistent physical object instances across views. This limitation is particularly critical in forensic settings, where evidence is tied to discrete objects and where scenes are often visually complex, unstructured and atypical in appearance. We present TRACE, a training-free framework for instance-level open-vocabulary inventory generation in 3D forensic evidence reconstruction. TRACE combines multiview instance formation, language-aligned aggregation, and 2D–3D semantic lifting within a unified Gaussian-based scene representation. It therefore moves beyond semantic scene querying towards a structured and language-addressable 3D inventory of physical evidence. The proposed formulation provides a bridge between investigator driven object assessment and instance-aware 3D scene understanding for forensic analysis.