Abstract
LiDAR super-resolution addresses the challenge of achieving high-quality 3Dperception from cost-effective, low-resolution sensors. While recenttransformer-based approaches like TULIP show promise, they remain limited tospatial-domain processing with restricted receptive fields. We introduce FLASH(Frequency-aware LiDAR Adaptive Super-resolution with Hierarchical fusion), anovel framework that overcomes these limitations through dual-domainprocessing. FLASH integrates two key innovations: (i) Frequency-Aware WindowAttention that combines local spatial attention with global frequency-domainanalysis via FFT, capturing both fine-grained geometry and periodic scanningpatterns at log-linear complexity. (ii) Adaptive Multi-Scale Fusion thatreplaces conventional skip connections with learned position-specific featureaggregation, enhanced by CBAM attention for dynamic feature selection.Extensive experiments on KITTI demonstrate that FLASH achieves state-of-the-artperformance across all evaluation metrics, surpassing even uncertainty-enhancedbaselines that require multiple forward passes. Notably, FLASH outperformsTULIP with Monte Carlo Dropout while maintaining single-pass efficiency, whichenables real-time deployment. The consistent superiority across all distanceranges validates that our dual-domain approach effectively handles uncertaintythrough architectural design rather than computationally expensive stochasticinference, making it practical for autonomous systems.