HGroupScene: Hierarchical Grouping and Similar Aggregation for 3D Semantic Scene Completion

Area of Science:

• Computer Vision
• Artificial Intelligence
• 3D Scene Understanding

Background:

• 3D Semantic Scene Completion (SSC) infers voxel occupancy and semantics from 2D data.
• Existing methods struggle with semantic interference and occlusion due to global attention or uniform voxel modeling.

Purpose of the Study:

• To propose HGroupScene, a unified framework for robust SSC.
• To integrate spatial priors and region-constrained reasoning to overcome limitations of current methods.

Main Methods:

• Introduced a Hierarchical Grouping Module for subregion partitioning and semantic aggregation using Gumbel-Softmax attention.
• Developed a dual-branch architecture with Explicit Constraint Branch (structural features) and Implicit Diffusion Branch (semantic reasoning).
• Implemented a Region-Constrained Feature Diffusion Mechanism for controlled feature propagation guided by structural priors.

Main Results:

• HGroupScene demonstrated competitive or superior performance on SemanticKITTI and SSCBench-KITTI360 datasets.
• Achieved strong results in both single-frame and multi-frame SSC settings.
• Validated the effectiveness of spatially structured semantic modeling.

Conclusions:

• HGroupScene effectively addresses semantic interference and occlusion in SSC.
• The proposed framework offers a robust approach to 3D semantic scene completion.
• Spatially structured semantic modeling is crucial for advancing SSC performance.