Budget-Constrained Generative Semantic Scene Completion for Airborne LiDAR via Selective Discrete Diffusion and State-Space Memory
LiDAR returns under limited computing resources. Dense voxel models waste computation on empty space, while discrete diffusion requires costly iterative sampling. We propose a budgetconstrained framework combining selective voxel processing, state-space memory, and crossmodal compensation. Budget-Aware Adaptive Voxel Ranking and Selection (BAVR) selects informative voxels under a compute budget. State-Space Memory-Augmented Directional Discrete Diffusion (SMD) captures spatial context and uses a distilled six-step denoising trajectory. Topological-Aligned Cross-Modal Dynamic Compensation (TCDC) fuses aerial imagery and Li-DAR using elevation constraints and confidence gating. On DALES, the method achieves 51.8%
Authors
- Ruohan Zong (ORCID: https://orcid.org/0000-0002-6499-3406)
- Chenrui Wei (ORCID: https://orcid.org/0009-0002-6963-0468)
- Mingzhen Liao
- Yichen Su
- Qirui Niu
Publication Details
- Published
- 2026-11-13
- DOI
- https://doi.org/10.22541/essoar.15012149/v1
- Primary Topic
- 3D Shape Modeling and Analysis
- Type
- preprint