Source-linked AI summary
Robust LIDAR Sensor-to-Sensor Domain Generalization via Accumulated-Point Scene Completion
Christoph B Rist, Larissa T Triess, Markus Enzweiler, Dariu M Gavrila
TL;DR
LIDAR sensor-specific sampling patterns hinder segmentation transfer, while domain generalization seeks to work on unseen sensors without their data or labels. The paper trains with occupancy supervision from accumulated source scans while keeping the input unchanged. It reports improved domain robustness at almost no source-accuracy cost, with outcomes dependent on the source–target pairing.
Problem
Sensor-specific and scene-specific LIDAR shifts complicate segmentation on unseen domains, where domain generalization uses no target scans or labels.
Source: §I · p. 1Method
The method adds occupancy prediction from accumulated source-domain points as an auxiliary training objective without altering the input scan.
Source: §I · p. 2Results
Accumulated-point supervision yields a domain-robustness benefit at almost no cost to source-domain accuracy, though outcomes depend on the source–target pairing.
Source: E. Discussion · p. 6Takeaways & Limitations
Richer geometric supervision on unchanged scans is a simple, stable lever for robustness when inter-domain semantics overlap sufficiently.
Source: E. Discussion · p. 6Effectiveness depends on the source–target pairing: transfers involving KITTI remain more difficult, and its validation split has limited diversity.
Source: E. Discussion · p. 6
Abstract
from arXiv · showhide
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.
I. INTRODUCTION
LIDAR domain generalization aims to support unseen sensor domains without access to their scans or labels, but sensor-specific shifts make transfer challenging. This work adds source-only occupancy supervision from accumulated points while leaving the input scan unchanged.
I. INTRODUCTION
- Domain generalization targets performance on unseen domains using only source data, avoiding the costly collection and annotation of data for every new domain.The paper motivates this setting by the diversity of weather, geography, traffic culture, and sensor generations faced by perception systems.Source: §I · p. 1
- Prior LIDAR domain-adaptation strategies can require sensor-specific priors, assumptions about future sensors, engineering effort, and may saturate across large discrepancies.The introduction contrasts these constraints with its input-preserving, source-only supervision scheme.Source: §I · pp. 1–2
- The study evaluates semantic segmentation across SemanticKITTI, nuScenes, and Waymo, which differ in LIDAR models and sampling patterns.The approach assumes source sequences can be registered reliably to generate aggregated supervision.Source: §I · p. 2
- The method adds a geometry-aware occupancy objective using accumulated source-domain points, changing supervision rather than the input LIDAR scan.Occupancy data is generated automatically; the auxiliary head can be discarded at test time, leaving semantic segmentation.Source: §I · p. 2
II. PREVIOUS WORK
The review distinguishes sensor-to-sensor and scene-to-scene LIDAR shifts, then surveys modality-agnostic feature learning and LIDAR-specific geometric approaches.
II. PREVIOUS WORK
- LIDAR domain shifts arise from sensor differences such as beam layout and range noise, as well as scene differences such as geography and traffic mix.The review groups prior approaches into generic feature learning and LIDAR-aware methods that exploit geometry or measurement physics.Source: §II · p. 2
- Modality-agnostic feature learning is broadly applicable, but LIDAR sparsity, class imbalance, and sensor-induced sampling artifacts motivate geometry- and surface-aware strategies.The review mentions completion, realistic re- and subsampling, and accumulation-based training, and points to surveys [14] and.Source: §II · p. 2
A. Entropy Minimization
The related-work section covers entropy-based adaptation and LIDAR methods that leverage geometry, sensor mapping, or mixed-sensor training.
A. Entropy Minimization
- Entropy minimization is a UDA baseline that encourages confident predictions on unlabeled target data without using sensor- or task-specific structure.Related image-domain work also uses auxiliary depth or geometry targets, while LIDAR pre-training on heterogeneous datasets is reported to improve robustness.Source: A. Entropy Minimization · p. 2, B. Domain Adaptation · p. 2
B. Domain Adaptation
- LIDAR domain-adaptation methods use surface or occupancy prediction, sensor-like resampling, or mixed-sensor pseudo-label training.Examples include canonical surface completion, occupancy pre-training,, implicit surface learning [8], resampling,,, and teacher-student mixing; some approaches use target data or incur computational costs.Source: B. Domain Adaptation · pp. 2–3
C. Domain Generalization
Prior LIDAR domain-generalization work varies source scan density or synthesizes target-like frames, while accumulated-point label propagation uses explicit geometric association.
C. Domain Generalization
- DGLSS simulates unseen sensors by subsampling a single source, but requires a sufficiently dense source and struggles in sparse-to-dense transfer.In that direction, the augment baseline and SCC-only variant outperform the full method.Source: C. Domain Generalization · p. 3
- Accumulated-sweep label propagation improves sensor robustness but relies on hand-designed geometric association rules and uses labeled accumulated points as model input.The paper contrasts this with its own use of a free-space class and joint occupied/free-space training without changing the input representation.Source: C. Domain Generalization · p. 3
- Other approaches strengthen robustness through source density augmentation or synthesize target-like frames using detailed sensor models.The latter avoids target data but requires accurate sensor specifications for ego-motion, spin distortions, and sampling patterns; synthetic-to-real indoor work also uses point-level pattern augmentation and instance mixing.Source: C. Domain Generalization · p. 3
D. Enriched Supervision Signal
The paper motivates enriching supervision without changing inputs as a route to better domain robustness. It trains with occupancy prediction using temporally accumulated scans, while segmentation still predicts point-wise labels from the input scan.
D. Enriched Supervision Signal
- The paper argues that richer supervision can improve domain robustness while keeping the input data unchanged.It situates this approach alongside self-supervised tasks, depth prediction, and privileged simulation depth.Source: D. Enriched Supervision Signal · p. 3
A. Overview
- The segmentation dataset pairs scans with point-wise labels and ego poses, and the model predicts labels from each scan.Source: A. Overview · pp. 3–4
- Ego-motion aligns successive sweeps into a denser static world frame, adding labeled surface points and sampled free-space points for occupancy supervision.At test time, the occupancy branch is ignored and the network performs standard point-cloud segmentation on source and unseen target scans.Source: A. Overview · p. 4
B. Training Data
Training supervision is built by fusing source scans into a shared world frame, filtering and thinning the points, and adding sampled free-space labels. Dynamic-object points from older frames are excluded to avoid motion trails.
B. Training Data
- Scans are registered using ego-motion and accumulated in a common world frame; nuScenes poses are further refined with local-window ICP.Points beyond 100 m are removed, and a 16 cm voxel grid thins dense regions while preferentially retaining labeled points.Source: B. Training Data · p. 4
- Points inside interpolated dynamic-object boxes are removed from non-current sweeps, while current-frame dynamic points are retained.This prevents motion trails in the fused world model while preserving instantaneous geometry.Source: B. Training Data · p. 4
- Free-space samples along registered rays are labeled non-occupied, adding a scene-completion signal alongside segmentation on accumulated surface points.Source: B. Training Data · p. 4
- The method uses raw 3D coordinates without reflectivity, domain-assumption-based subsampling, or source-sweep augmentation.Temporal accumulation increases the number of semantically labeled points compared with single sweeps.Source: B. Training Data · p. 4
C. Model
The model voxelizes raw 3D points into pillars, processes them with a multi-resolution backbone, and queries a Coordinate MLP head at arbitrary locations. The same head supports accumulated-surface and free-space supervision during training, while inference queries only LIDAR points.
C. Model
- Raw 3D points are voxelized into point pillars and processed by a multi-resolution backbone before a Coordinate MLP head produces predictions at arbitrary 3D locations.Source: C. Model · p. 4
- Training can query the shared output head at accumulated surface points and sampled free-space locations without changing the output head.Alternatively, voxel-grid supervision or backprojected grid outputs can provide equivalent supervision.Source: C. Model · p. 4
- At test time, the free-space class is dropped, semantic probabilities are normalized, and predictions are queried only at LIDAR points.Source: C. Model · p. 4
A. Existing Approaches and Label Sets
Across the evaluated label mappings, accumulated-point supervision improves some unseen-domain transfers, with outcomes depending on the dataset pair and label set. The datasets differ in sensor configurations and sampling patterns, and evaluation uses mapped 10-class label sets.
A. Existing Approaches and Label Sets
- The comparison includes C&L, which uses unlabeled target data, DGLSS, which subsamples inputs to simulate sensors, and 3DLabelProp, which uses accumulated points as input and supervision.C&L and DGLSS use MinkowskiNet-based architectures and have source-domain performance comparable to the proposed method.Source: A. Existing Approaches and Label Sets · p. 4
- The datasets map their taxonomies to two 10-class variants: 10-A and 10-B, which differ in how terrain, vegetation, sidewalk, and walkable classes are grouped.Source: A. Existing Approaches and Label Sets · p. 4
B. Dataset Characteristics
- KITTI uses a single 64-beam sensor, nuScenes uses a 32-beam sensor and five recent scans as input, and Waymo coordinates multiple roof-mounted LIDARs.The configurations produce distinct sampling geometries and point densities.Source: B. Dataset Characteristics · p. 5, Figure 3 · p. 5
C. Single-Source Domain Generalization
- On 10-B, nuScenes- and Waymo-trained models improve transfers to the other dataset by 8% and 7%, while transfers involving KITTI slightly degrade against the segmentation baseline.Source performance is nearly unchanged except for a slight KITTI drop; the method modestly beats DGLSS on Waymo-to-nuScenes, but not on KITTI transfers.Source: C. Single-Source Domain Generalization · p. 5
B. Dataset Characteristics
- Table I organizes training-split scene counts and annotated LIDAR-point counts for individual datasets and pairwise combinations.Source: Table I · p. 6
C. Single-Source Domain Generalization
- On 10-A, accumulated-point supervision improves transfers involving KITTI more than on 10-B, and 3DLabelProp has higher domain-generalization performance overall.For KITTI-to-nuScenes, accumulated-point training can surpass C&L relative to the segmentation baseline.Source: C. Single-Source Domain Generalization · p. 6
D. Ablations on Single-Source and Multi-Source Domain Generalization
Occupancy supervision leaves source-domain performance very similar, while multi-source training has mixed effects on transfer. Accumulated-points training often outperforms the corresponding segmentation baseline, but not in every source–target pairing.
D. Ablations on Single-Source and Multi-Source Domain Generalization
- Source-domain performance remains very similar with and without occupancy supervision, while all-source training slightly reduces per-dataset source accuracy but exceeds the reported single-source domain-generalization results.Adding KITTI nearly closes the segmentation baseline’s gap to its best single-source domain-generalization performance.Source: D. Ablations on Single-Source and… · p. 6
- Accumulated-points training outperforms the corresponding segmentation baselines for KITTI+nuScenes and KITTI+Waymo, but gains are modest or largely neutral relative to single-source training.For transfer to KITTI, nuScenes+Waymo gives a clear boost over single-source training, but the segmentation baseline scores better than accumulated-points training.Source: D. Ablations on Single-Source and… · p. 6
- Multi-source combinations do not uniformly improve transfer: KITTI+Waymo gains are modest, KITTI+nuScenes is largely neutral, and nuScenes+Waymo benefits transfer to KITTI.The nuScenes+Waymo segmentation baseline attains the better score for transfer to KITTI.Source: D. Ablations on Single-Source and… · p. 6
E. Discussion
Accumulated-points supervision offers a domain-robustness benefit with negligible source-domain trade-off, and the benefit varies by dataset pairing. The discussion attributes potential transfer gains to geometry-aware features, while noting that simple data pooling is not a general solution.
E. Discussion
- The source-domain accuracy trade-off is negligible, and accumulated-points targets may encourage geometry- and boundary-aware features that generalize across sampling patterns.Because inputs remain unchanged, the comparison isolates richer supervision from input transformations.Source: E. Discussion · p. 6
- The benefit depends on source–target pairing: Waymo↔nuScenes transfers are pronounced, whereas KITTI transfers remain more difficult.The authors suggest broader dataset gaps may matter and note that limited diversity in KITTI’s validation split may skew cross-domain results.Source: E. Discussion · p. 6
- Simple multi-source pooling does not solve domain generalization; accumulated-points supervision generally helps, with modest gains, when inter-domain semantics overlap sufficiently.Scene composition, geography, object frequency and scale, and static-versus-dynamic class balance may reduce the utility of features learned in another domain.Source: E. Discussion · p. 6
V. CONCLUSION
The paper evaluates unchanged-input, scene-completion supervision for LIDAR domain generalization across KITTI, nuScenes, and Waymo. Results depend strongly on dataset pairing: Waymo–nuScenes transfers are substantially stronger than transfers involving KITTI.
V. CONCLUSION
- The method keeps the input scan unchanged and enriches supervision with scene-completion targets from accumulated points, evaluated across combinations of KITTI, nuScenes, and Waymo.The authors state that prior work had not isolated this effect without modifying the input.Source: §V · p. 6
- Waymo–nuScenes transfers generalize substantially better than transfers involving KITTI, which remain challenging.The approach is competitive with a classic surface-based domain-adaptation baseline and a more recent domain-generalization approach, but does not surpass the latest methods.Source: §V · p. 7
- Input-preserving supervision is complementary to input-altering strategies, which the authors say can be combined to aggregate their strengths.The conclusion describes the supervision change as capable of considerably shifting domain-generalization outcomes.Source: §V · pp. 6–7
- The proposed next step adapts the setup to unsupervised domain adaptation by deriving target-domain completion labels for auxiliary training supervision.The proposal keeps single-scan inference unchanged and tests whether target-side completion can close the remaining gap.Source: §V · p. 7