Planar structures, ubiquitous in man-made indoor environments, enable compact and accurate scene decomposition. Recent methods distill planar features into learning-based MVS geometries for multi-view 3D plane estimation. However, the implicit nature of planar features hinders the semantic–geometry alignment within individual planar instances, resulting in distorted geometry and fragmented semantics. To address this, we propose PIGS, the first Planar-Instance Gaussian Splatting framework that explicitly targets planar instances and enforces planar semantic–geometry alignment through planar structural priors, free from feature distillation. Starting from planar-unaware geometry, our method first employs the GHPS module to achieve annotation-free single-view planar segmentation, then utilizes the MVSA module to perform multi-view instance association and planarity aggregation, and finally leverages the PIGO module to initialize and optimize Planar-Instance Gaussian primitives through planar-aware loss functions during instance-level splatting. Followed by a planar-aware ball-pivoting strategy for final plane mesh extraction, PIGS produces semantically and geometrically aligned 3D planes, closely matching RGB references. Extensive experiments on hundreds of indoor scenes and comprehensive ablations demonstrate the superior performance, while module-agnostic evaluations and outdoor experiments validate its generalization capability.
