Compose, Route, Complete: Text-Routed Vision Prototypes for Multimodal Point Cloud Completion
Abstract
Point cloud completion aims to recover complete 3D shapes from partial observations. Recent methods introduce single-view images as auxiliary guidance, yet visual features are bound to the camera viewpoint and carry little information about occluded regions. Text generated by a Vision-Language Model can complement images with view-invariant structural priors. However, treating such text as just another feature source and fusing it directly with image features discards its compositional structure and even degrades quality. We argue that the true value of text lies in decomposing an object into semantic parts, and propose CRCNet, a multimodal framework built on a compose-and-route paradigm. Each part phrase selects relevant visual features via cross-attention to form a concept prototype, and each coarse point retrieves prototypes by its affinity to the part phrases for targeted refinement. We further introduce a Geometry-Gated Semantic Dispatch (GGSD) module: while routing decides which prototypes each point attends to, a proximity-aware gate controls how much guidance it receives, based on its geometric displacement from the observed surface. Since each point is explicitly routed through part-level phrases, our framework is highly interpretable. Experiments on ShapeNet-ViPC show that CRCNet achieves state-of-the-art performance and generalizes strongly to unseen categories and to real-world LiDAR scans.