Driving · 2026-10-30
Robust LIDAR Sensor-to-Sensor Domain Generalization via Accumulated-Point Scene Completion
Christoph B Rist, Larissa T Triess, Markus Enzweiler, Dariu M Gavrila
AI-assisted summary based on the paper. Consult the linked source for methods and limitations. Benchmark numbers appear after editor verification.
TL;DR
The authors propose a source-only domain generalization method for LiDAR semantic segmentation that keeps the input scan unchanged. During training, they aggregate source-domain scans into an accumulated world-frame “world model” and use it as enriched supervision via an occupancy prediction auxiliary task (including free-space points). At test time, they drop the occupancy head for pure segmentation or keep it for scene completion.
Why it matters
Sensor-specific sampling patterns can prevent LiDAR segmentation models from transferring to completely unseen sensors. The paper targets this sensor-to-sensor domain generalization setting using only source data and labels, improving generalization through supervision enrichment rather than input modification or explicit domain adaptation/refinement.
Method
- Train on the source domain with standard semantic supervision plus an occupancy auxiliary task generated from accumulated multi-scan points in a common world frame (including free-space points along registered rays).
- At test time, use the same single-frame LiDAR input as during training, and ignore the occupancy branch (or keep it for scene completion) with no additional inference cost.
- Build the occupancy supervision by registering successive sweeps using ego-motion (and ICP pose refinement for nuScenes), removing dynamic-object points from non-current frames, and sampling free-space as non-occupied points.
Abstract (from arXiv)
We study domain generalization for LIDAR semantic segmentation. Deep neural networks achieve strong performance on 3D point clouds, yet real-world LIDAR exhibits sensor-specific sampling patterns that hinder transfer across datasets. In many applications, performance on completely unseen sensors is crucial. A model should preferably work on a new setup without prior access to its data or labels. To this end, we introduce a source-only training scheme that adds an occupancy prediction auxiliary objective. We aggregate source-domain scans into a simple world model and use it as the supervision signal, replacing single-scan standard segmentation training. A key property of our approach is that it does not alter the input scan. During training and testing the network consumes the same single-frame LIDAR input, only the supervision signal is enriched via the occupancy objective. At deployment, there is no runtime overhead as the auxiliary head can be dropped for pure segmentation, or kept to obtain scene completion alongside semantic segmentation. In an experimental evaluation across distinct LIDAR datasets, we demonstrate that a model's generalization to new domains can be significantly affected by how it is supervised. In certain setups, performance on unseen domains improves without any explicit domain adaptation (no target scans, no target labels, no sensor priors) and without prediction refinement. At the same time, our method maintains competitive source-domain segmentation performance and incurs no additional inference cost.
Related papers
- Into the LidarVerse: A Real-Time Unified LiDAR World Model Across Tasks and Sensors
- Let LiDAR Decide Where to Look: Geometry-Measured Token Selection for Efficient Sparse Latent World Models
- LO-World: Unified BEV Latent Modeling for Joint 4D Occupancy and LiDAR Forecasting
- CRISP: Fixing Flying Pixels in Latent LiDAR Generation via Diffusion Decoding
- Good Views Are Not Enough: Global Counterexamples for Gaussian-Regularized JEPAs
Publication record
Bibliographic evidence describes where and when a record appeared. It does not validate scientific claims.
First public release: 2026-10-29 · Indexed by Pub-AI: 2026-10-30 UTC
Robust LIDAR Sensor-to-Sensor Domain Generalization via Accumulated-Point Scene Completion
Preprint posted: 2026-10-29 · Original record ↗
- arxiv
- 2610.39516
Explicit source versions (1)
- v1 · 2026-10-29
- Preprint available
Source date: 2026-10-29 · First detected: 2026-10-30 UTC · Historical baseline
Source evidence (2)
- Official evidence ↗ · RSS guid=oai:arXiv.org:2610.39516v1; arxiv:announce_type=new; item pubDate (official metadata)
- Official evidence ↗ · atom:published (official metadata)
Abstract · arxiv
We study domain generalization for LIDAR semantic segmentation. Deep neural networks achieve strong performance on 3D point clouds, yet real-world LIDAR exhibits sensor-specific sampling patterns that hinder transfer across datasets. In many applications, performance on completely unseen sensors is crucial. A model should preferably work on a new setup without prior access to its data or labels. To this end, we introduce a source-only training scheme that adds an occupancy prediction auxiliary objective. We aggregate source-domain scans into a simple world model and use it as the supervision signal, replacing single-scan standard segmentation training. A key property of our approach is that it does not alter the input scan. During training and testing the network consumes the same single-frame LIDAR input, only the supervision signal is enriched via the occupancy objective. At deployment, there is no runtime overhead as the auxiliary head can be dropped for pure segmentation, or kept to obtain scene completion alongside semantic segmentation. In an experimental evaluation across distinct LIDAR datasets, we demonstrate that a model's generalization to new domains can be significantly affected by how it is supervised. In certain setups, performance on unseen domains improves without any explicit domain adaptation (no target scans, no target labels, no sensor priors) and without prediction refinement. At the same time, our method maintains competitive source-domain segmentation performance and incurs no additional inference cost.
