Following the validation of geometric features on object parts, we scaled our approach to full 3D scenes using a query-based 3D segmentation model designed to handle the complexity and scale of scene-level data. Unlike the point-wise classification employed in our preliminary experiments, this model predicts masks over "superpoints"—precomputed over-segments of the scene—which significantly reduces computational overhead while preserving boundary adherence. The workflow begins with Voxelization and Feature Extraction, where the scene is discretized into sparse voxels carrying 6D features (XYZ + RGB). A lightweight backbone, such as a Minkowski sparse-convolution stack or an MLP, processes these voxels into high-dimensional embeddings, which are then pooled via scatter-mean operations into compact superpoint features.

The core of the architecture lies in its Transformer Decoder, which treats segmentation as a direct set prediction problem. The model initializes a combined set of learnable queries: K Instance Queries, sampled from superpoint features with learned biases to target potential objects, and C Semantic Queries, representing learned class embeddings. These concatenated queries pass through a 6-layer decoder where they first interact globally via self-attention to resolve context and occlusions, and then attend to the superpoint tokens via cross-attention to extract localized mask information. Finally, the decoded queries branch into two prediction heads: a linear classifier that assigns a semantic class (or "no-object") to each query, and a mask head that generates segmentation masks via a dot product between the query vectors and the superpoint features. This architecture allows the system to effectively "ask questions" about the scene, robustly identifying and isolating individual object instances even in cluttered environments.