Paper accepted at NeurIPS 2026: “Towards Real-Time Full-Waveform LiDAR Transformers”
Full-waveform LiDAR (FWL) records the complete temporal profile of the returned light intensity rather than only peak distances, providing cues that ordinary point clouds lack, such as those needed to distinguish multipath reflections from glass and other reflective surfaces. The Yoshioka group presented Ghost-FWL, a large-scale FWL dataset for ghost detection and removal, at CVPR 2026. Because FWL holds a temporal waveform for every pixel, however, its data volume is large, and processing it in real time with large models such as Transformers remains a challenge.
This paper works toward real-time FWL Transformers through intensity-guided token reduction, which reduces the tokens a Transformer must process based on the received intensity, and physics-aware augmentation, which takes the physical characteristics of LiDAR into account. The work is a collaboration with the Isogawa Lab at Keio University. It will be presented at NeurIPS 2026, held in December 2026 across three cities: Sydney, Atlanta, and Paris.
[Paper]
Kotaro Oishi*, Kazuma Ikeda*, Ryosei Hara, Ryo Yoshida, Mariko Isogawa, Kentaro Yoshioka, “Towards Real-Time Full-Waveform LiDAR Transformers via Intensity-Guided Token Reduction and Physics-Aware Augmentation,” The Fortieth Annual Conference on Neural Information Processing Systems (NeurIPS 2026), December 2026. (* co-first authors)