E3CoverNet: Progressive E(3)-Aware 3D Backbone for Geometry-Sensitive Vision-Language Alignment
Abstract
3D vision-language understanding is central to multimedia research. Representative tasks such as visual grounding and text-to-shape retrieval require aligning 3D geometry with natural language that often depends on fine-grained geometric cues. Most existing 3D-language models build on generic backbones with limited geometric inductive bias, leaving them brittle to spatial transformations. Conversely, strictly invariant encoders risk collapsing the very spatial distinctions that geometry-sensitive queries depend on. We introduce Equivariant Covering Networks on E(3) (E3CoverNet), a 3D backbone designed to close this gap. Rather than directly modeling the full E(3) transformation space, E3CoverNet progressively builds equivariant representations along the subgroup chain E(1) ⊂ E(2) ⊂ E(3). At each stage, a learned covering set discretizes the corresponding symmetry group, and a geometry-aware attention layer synchronously updates invariant point features and equivariant coordinates through a dual-path design. We plug E3CoverNet into standard 3D-language pipelines and evaluate it on text-to-shape retrieval and 3D visual grounding benchmarks, where it outperforms competitive baselines across both tasks. A geometry-sensitive diagnostic further shows that E3CoverNet substantially improves robustness to spatial perturbations without sacrificing the semantic discrimination required for complex spatial queries.