UniFlow: Zero-Shot LiDAR Scene Flow for Autonomous Driving

Conference: ECCV 2026
Paper: CVF Open Access
OpenReview: https://eccv.ecva.net/virtual/2026/poster/3875
Code: https://github.com/KTH-RPL/OpenSceneFlow
Area: Autonomous Driving
Keywords: LiDAR Scene Flow / Zero-Shot Generalization / Multi-Dataset Training / Velocity Domain Gap / Geometric Data Augmentation

TL;DR

UniFlow overturns the long-standing belief from 3D detection and segmentation that multi-dataset LiDAR training degrades performance, proving that class-agnostic scene flow models transfer robustly across sensors when trained on unified datasets with temporal alignment and minimal geometric augmentations, achieving state-of-the-art results on Waymo and nuScenes while outperforming dataset-specific models on unseen TruckScenes and AEVAScenes benchmarks.

Background & Motivation

In 3D autonomous driving perception, LiDAR scene flow estimation aims to determine the 3D displacement vector for every point between consecutive point cloud sweeps, providing a critical geometric primitive for tracking dynamic obstacles, path planning, and 4D world modeling. Contemporary state-of-the-art feedforward scene flow models (such as SSF, Flow4D, and \(\Delta\)Flow) have achieved centimeter-level accuracy, but almost exclusively train and evaluate on single datasets in isolation. Autonomous driving datasets exhibit significant physical discrepancies in sensor hardware—Argoverse 2 uses two asynchronous 32-beam sensors, nuScenes uses a single 32-beam sensor, and Waymo relies on a custom 64-beam sensor. Because of this stark heterogeneity in point densities and mount geometries, researchers have traditionally assumed that multi-dataset training is prone to severe domain conflict.

This assumption was largely inherited from higher-level 3D vision tasks like 3D object detection and semantic segmentation, where naively pooling multiple datasets notoriously degrades accuracy relative to single-dataset specialists. However, the core tension in semantic perception stems from conflicting, human-defined taxonomy boundaries—for example, nuScenes excludes riders from the bicycle bounding box while Waymo includes them, creating direct label contradictions across domains. In contrast, scene flow is inherently class-agnostic low-level motion estimation governed by the physical laws of rigid and non-rigid motion in 4D space, making it fundamentally less vulnerable to semantic label ambiguity.

Empirical cross-domain analysis reveals that existing single-dataset models already exhibit nascent generalization capabilities, and the primary bottleneck is not point density disparity but rather an uneven distribution of object velocities across datasets (e.g., suburban highway vs. urban downtown). Core idea: train feedforward LiDAR scene flow architectures across unified multi-source driving datasets without architectural modifications, standardizing temporal frame intervals and applying minimal geometric augmentations to learn generalizable 4D motion priors that transfer zero-shot across diverse sensors and driving scenarios.

Method

Overall Architecture

UniFlow establishes a scalable multi-dataset training paradigm for feedforward LiDAR scene flow. Given two successive unorganized point clouds \(P_t, P_{t+1}\) captured at consecutive timestamps, the model directly estimates the per-point 3D motion vector field \(F_t \in \mathbb{R}^{N \times 3}\). The pipeline integrates multi-dataset temporal standardization, ground removal, simple physics-grounded geometric augmentations, and off-the-shelf feedforward architectures.

%%{init: {'flowchart': {'rankSpacing': 24, 'nodeSpacing': 28, 'padding': 6, 'wrappingWidth': 400}}}%%
flowchart TD
    A["Multi-Dataset LiDAR Inputs<br/>AV2 / Waymo / nuScenes"] --> B["Dataset Standardization & Temporal Alignment<br/>Standardize to 10 Hz + LineFit ground removal"]
    B --> C["Cross-Domain Geometric Augmentation<br/>Height randomization + LiDAR beam dropping"]
    C --> D["Feedforward Scene Flow Backbone<br/>Unmodified SSF / Flow4D-XL / ΔFlow"]
    D --> E["Zero-Shot Cross-Sensor & Long-Range Evaluation<br/>TruckScenes highway truck / AEVAScenes FMCW"]

Key Designs

1. Dataset Standardization & Temporal Alignment: Reconciling Motion Parameterization Scene flow architectures explicitly parameterize motion as the spatial displacement vector between consecutive frames. If raw sensor capture frequencies differ, objects traveling at the exact same physical velocity produce vastly different point displacements, confounding geometric learning. To eliminate this temporal domain gap, UniFlow standardizes LiDAR frame rates across datasets to 10 Hz by down-sampling 20 Hz nuScenes streams to match AV2 and Waymo. Furthermore, to address nuScenes' sparse 2 Hz track annotations without introducing motion interpolation artifacts or label noise, the training pipeline exclusively pairs annotated sweeps with their true temporal neighbors. Non-informative ground returns are stripped across all datasets using LineFit, and ego-vehicle motion compensation is uniformly applied.

2. Cross-Domain Geometric Augmentation: Bridging Height and Density Gaps To bridge domain gaps across distinct vehicle platforms and beam configurations without complex generative networks, UniFlow introduces two targeted physical augmentations. First, to bridge the viewpoint elevation gap between passenger sedans and heavy commercial trucks whose sensors sit meters higher, the pipeline adds random vertical translation (\(z \pm 0.5 \sim 2.0\,\text{m}\) with probability \(p=0.8\)). Second, to prevent models from overfitting to dense beam structures (such as Waymo's 64-beam array) when deploying to sparse sensors (such as nuScenes' 32-beam setup), it applies random alternating beam dropout (ray dropping on odd/even scan lines with \(p=0.35\)). Ray dropping is omitted on nuScenes due to its pre-existing sparsity.

3. Velocity Distribution Scaling & Model Expansion: Mitigating Motion Extrapolation Cross-domain speed bucket analyses show that models struggle when encountering velocity ranges unrepresented in their training data. By training on a diverse mixture of AV2, Waymo, and nuScenes, UniFlow broadens the velocity spectrum, enabling robust prediction on fast-moving objects. To capitalize on the increased data scale without underfitting, the authors scale up Flow4D into Flow4D-XL by expanding channel capacities in its 4D spatio-temporal convolutional backbone. Furthermore, because voxelized scene flow backbones operate with invariant local voxel resolutions, the model seamlessly scales to long-range inference (extending from 35 m to beyond 100 m) by merely expanding the dynamic voxel grid bounds at test time without retraining.

Loss & Training

UniFlow trains on the pooled dataset mixture using a standard supervised point-wise displacement loss combined with spatial smoothness regularizers, requiring no dataset frequency re-weighting or custom losses: $\(\mathcal{L} = \frac{1}{|P_{fg}|} \sum_{p_i \in P_{fg}} \|\hat{f}_i - f_i^*\|_1 + \lambda_{smooth} \mathcal{L}_{smooth}\)$ Here \(P_{fg}\) denotes foreground dynamic and static points, and \(\hat{f}_i, f_i^*\) are predicted and ground-truth 3D flow vectors. Multi-task experiments incorporating a shared semantic classification head reveal that semantic supervision severely degrades out-of-distribution transfer (due to label ambiguity across datasets) while offering negligible benefit to motion estimation; hence, pure motion supervision is preserved.

Key Experimental Results

Main Results

UniFlow is benchmarked against supervised and optimization-based methods on Argoverse 2, Waymo, and nuScenes, as well as zero-shot evaluated on unseen TruckScenes (heavy commercial trucks on highways) and AEVAScenes (continuous-wave FMCW sensors). Primary metrics are Three-way End-Point-Error (Three-way EPE, in cm) and Dynamic Bucket-Normalized Mean EPE.

Table 1: Cross-Dataset Performance & Zero-Shot Generalization (Excerpts from Tables 1, 4, 6)

Benchmark / Evaluation Split	Model & Training Setup	Foreground Dynamic FD (cm) ↓	Dyn. Mean EPE ↓	Fast Speed [2.0, \(\infty\)) EPE ↓	Relative Gain
Waymo Val	Flow4D (Waymo Single-Domain)	4.58	0.215	0.130	Baseline
Waymo Val	Flow4D-XL (UniFlow, Ours)	3.83	0.191	0.110	-11.2%
Waymo Val	\(\Delta\)Flow (Waymo Single-Domain)	3.27	0.198	0.114	Baseline
Waymo Val	\(\Delta\)Flow (UniFlow, Ours)	3.07	0.188	0.096	-5.1%
TruckScenes (Zero-Shot)	Flow4D (TruckScenes In-Domain)	44.87	0.456	0.413	In-Domain Base
TruckScenes (Zero-Shot)	Flow4D (Waymo Single-Domain)	116.01	0.336	0.586	Out-of-Domain Base
TruckScenes (Zero-Shot)	Flow4D-XL (UniFlow, Ours)	68.41	0.281	0.389	-38.4% (vs in-domain)
TruckScenes (Zero-Shot)	\(\Delta\)Flow (UniFlow, Ours)	45.76	0.283	0.226	-29.6% (vs in-domain)

Table 2: Zero-Shot Generalization on AEVAScenes FMCW LiDAR (Excerpt from Table 7)

Method Backbone	Training Source Domain	Three-way Mean (cm) ↓	Foreground Dynamic FD (cm) ↓	Dyn. Bucket-Norm. Mean ↓	Car Category EPE ↓
SSF	Waymo Single Dataset	20.59	59.38	0.811	0.513
SSF	UniFlow (Ours)	10.80	29.76	0.544	0.239
Flow4D	Waymo Single Dataset	8.14	18.89	0.408	0.152
Flow4D-XL	UniFlow (Ours)	6.14	13.16	0.338	0.114
\(\Delta\)Flow	Waymo Single Dataset	7.01	16.27	0.391	0.123
\(\Delta\)Flow	UniFlow (Ours)	6.65	13.17	0.303	0.112

Ablation Study

A step-by-step ablation on TruckScenes zero-shot performance reveals how dataset mixture, geometric augmentations, and model capacity contribute to out-of-distribution transfer (Flow4D backbone, Table 10):

Stage Configuration	FD EPE (cm) ↓	Dyn. Normalized EPE ↓	Slow [0, 0.5)	Mid [0.5, 1.0)	Fast [2.0, \(\infty\))	Note
Flow4D (Waymo Single-Domain)	116.01	0.336	0.071	0.129	0.586	Lacks viewpoint height and speed diversity
+ Unified Dataset Mixture	93.76	0.310	0.040	0.116	0.594	Multi-dataset scale reduces FD by 22.25 cm
+ Geometric Augmentation	68.32	0.301	0.047	0.111	0.407	Height shift cuts fast-speed error dramatically
+ XL Backbone Scaling	68.41	0.281	0.033	0.097	0.389	Greater capacity unleashes multi-dataset gains

Key Findings

• Velocity diversity is the primary driver of generalization: In single-dataset evaluations, Flow4D trained on Waymo outperforms Flow4D trained natively on AV2 when tested on fast-moving objects in AV2 (0.098 vs 0.103 EPE), directly reflecting Waymo's richer distribution of higher-speed trajectories.
• Semantic heads impede cross-domain motion generalization: Adding a 4-class semantic head causes TruckScenes out-of-domain semantic mIoU to collapse from 64.6% to 40.5% (Table 9) due to cross-dataset taxonomy misalignment, demonstrating why low-level geometric training should remain decoupled from high-level semantics.
• Near-range spatial sparsity emulates long-range geometry: Models trained only on points within \(0 \sim 35\,\text{m}\) achieve a 40.6% error reduction when evaluated out to \(100\,\text{m}\) on TruckScenes without retraining (Table 8), proving that angular beam sparsity nearby serves as a natural proxy for distant physical returns.

Highlights & Insights

• Challenging conventional wisdom with minimal intervention: Instead of designing heavy domain-alignment architectures, UniFlow shows that standard scene flow backbones excel out-of-the-box when fed diverse data, highlighting that low-level vision lacks the taxonomy bottlenecks of semantic tasks.
• Velocity gap as a critical 4D domain axis: Prior domain adaptation work focused almost entirely on spatial geometry (density, beam pattern), while UniFlow highlights that temporal velocity coverage is equally critical for time-dependent perception.
• Cost-effective physical augmentations: Height shifts and beam dropouts provide strong transfer leverage to unseen vehicle form factors (trucks) and sensor modalities (FMCW) at near-zero computational overhead.

Limitations & Future Work

• Rolling shutter timing remains unmodeled: Mechanical spinning, MEMS mirrors, and solid-state FMCW sensors have distinct intra-frame rolling shutter latencies that are not explicitly modeled beyond 10 Hz frame rate downsampling.
• Sensitivity to ego-motion estimation: UniFlow relies on accurate dataset-provided ego-vehicle odometry for background compensation; degradation in localization poses could impair foreground motion isolation.
• Future directions: Integrating self-supervised pre-training on unlabeled multi-sensor streams and extending the pipeline to hostile weather artifacts like heavy rain and fog.

Related Work & Insights

• vs Optimization-based Scene Flow (NSFP, FastNSF, EulerFlow): Per-scene optimization methods require tens of minutes to 24+ hours per sequence to converge; UniFlow delivers comparable or superior accuracy in feedforward real-time inference (\(<50\,\text{ms}\)).
• vs Single-domain Feedforward Estimators (SSF, Flow4D, \(\Delta\)Flow): UniFlow serves as an orthogonal, general training paradigm that readily boosts any standard feedforward scene flow backbone across in-domain and out-of-domain benchmarks.

Rating

• Novelty: ⭐⭐⭐⭐☆ Reframes LiDAR cross-sensor generalization around velocity distributions and class-agnostic motion, dismantling a long-held domain adaptation assumption.
• Experimental Thoroughness: ⭐⭐⭐⭐⭐ Rigorous validation across 5 distinct datasets, encompassing cross-sensor, cross-chassis, long-range, frame-rate, and multi-task ablations.
• Writing Quality: ⭐⭐⭐⭐⭐ Exceptionally structured, insightful, and transparent about empirical results and failure modes.
• Value: ⭐⭐⭐⭐⭐ Offers a foundational, practical recipe for training foundation motion models in autonomous driving and 4D scene understanding.

Related Papers

• [ECCV 2026] FrozenDrive: Zero-Shot Text-Guided Driving Scene Generation and Data Augmentation
• [ECCV 2026] DriveVA: Video Action Models are Zero-Shot Drivers
• [ECCV 2026] SynFlow: Scaling Up LiDAR Scene Flow Estimation with Synthetic Data
• [ECCV 2026] LiFlow: Flow Matching for 3D LiDAR Scene Completion
• [CVPR 2025] Zero-Shot 4D Lidar Panoptic Segmentation