Hierarchical Scoring With 3D Gaussian Splatting for Instance Image-Goal Navigation

Abstract

Instance Image-Goal Navigation (IIN) asks an agent to locate the specific object instance shown in a goal image. Existing 3D Gaussian Splatting (3DGS) based methods rely on pose-centric search—sampling many viewpoints, rendering them, and comparing against the goal—which is inefficient in continuous 6-DoF space. We instead treat the 3DGS scene as a joint semantic-geometric ray field and reformulate localization as inference over this field. Our framework, GauScoreMap, performs hierarchical scoring: a training-free global stage produces a CLIP relevance field over Gaussians and prunes the scene to a small set of candidate regions, and a local stage performs ray-image cross-attention within those regions and triangulates the 6D camera pose from the top-scoring rays. On the HM3D IIN benchmark, GauScoreMap sets a new state of the art with more than an order of magnitude lower per-query latency than the prior 3DGS-based work, and we validate the pipeline in a real-world environment.