GAMPoint: Geometry-augmented vision transformer with a multimodal-ready interface for 3D point-cloud segmentation

Various 3D point cloud semantic segmentation methods have been proposed, yet they remain challenged by limited use of geometric cues and inaccurate predictions around thin structures, object boundaries, and rare classes in cluttered indoor scenes. To address these limitations, we propose GAMPoint, a transformer-based framework for 3D semantic segmentation. Specifically, Enhanced Relative Positional Encoding (EnhancedRPE) augments relative positional modeling with local surface-variation cues to strengthen geometry-aware feature learning. In addition, Conditional Multimodal Gating (CMG) is introduced as an input-side embedding interface. In GAMPoint, XYZ coordinates provide the basic point-cloud geometry, while CMG incorporates available auxiliary attributes, such as RGB colors and surface normals, before the serialized encoder. For RGB-D reconstructed datasets such as ScanNet, XYZ coordinates already contain depth-derived geometric information; raw depth is not used as an additional per-point channel. GAMPoint incorporates EnhancedRPE and CMG into Point Transformer V3 and is evaluated on S3DIS and ScanNet. Experimental results show competitive performance and modest gains over strong point-transformer baselines. Class-wise and repeated-run analyses further indicate category-dependent improvements and consistent gains across matched repeated runs under the reported settings. Code is available at github.com/airmanss/ GAMPoint.