Title: Simple Enhancements to Elevate Lidar Semantic Scene Completion

Original: arXiv:2606.03992v1 Announcement Type: New Abstract: This study explores "free lunch" methods to enhance lidar semantic scene completion (SSC) performance without necessitating intricate architectural modifications. Initially, we show that providing input point clouds with semantic pseudo-labels derived from readily available segmentors markedly boosts the capabilities of current models. Through a comparison with an oracle, we confirm that robust semantic priors are the main factor behind improvements in mean Intersection over Union (mIoU). Additionally, we incorporate visibility data into the input lidar scan to differentiate between empty and unknown regions, yielding a supplementary performance increase across the evaluated architectures. Leveraging these straightforward improvements, we find that legacy models can stay competitive with cutting-edge systems and even surpass them. Our code is accessible at https://github.com/astra-vision/SSC-Priors.

Source: arXiv Generated at: 2026-06-03 00:00:00 UTC