LiDAR scene flow estimates point-wise motion between two consecutive scans, referred to as the source and target. Leading self-supervised methods typically minimize the Chamfer loss, the nearest neighbor distance between the flow-compensated source and the target. However, nearest-neighbor search does not enforce motion rigidity, often leading to inconsistent flows within object instances. Existing approaches address this issue with additional regularization terms, but flow consistency among points remains limited, especially for large objects. We propose RVLoss, a self-supervised loss that incorporates motion rigidity by design through a runoff vote mechanism. Our key observation is that the point-wise motion, calculated from nearest neighbor search, can often be grouped into a small set of dominant flow candidates by voting (top-k voting). Furthermore, when compensating the source by these candidates, the flow that best represents the underlying rigid motion often yields the highest consensus after a second voting (top-1 voting). Based on this insight, we incorporate the two-stage runoff vote into loss design and create cluster-wise rigid flows and free-form flows as pseudo-labels for self-supervised learning. RVLoss can be seamlessly integrated into existing feedforward architectures. Experiments on the Argoverse2 2026 Challenge show that models trained with RVLoss achieve state-of-the-art performance among self-supervised approaches, outperforming baseline models trained with alternative loss designs by 20%. Moreover, cross-dataset evaluations demonstrate consistent performance improvements across four additional datasets. Code will be released upon acceptance.
Figures & tables
Figure 1: Overview of a runoff vote. a) A car moves rigidly and thus exhibits consistent flows among points. b) The widely used Chamfer distance breaks motion rigidity due to the local nearest neighbor search. As a result, multiple flow candidates emerge within the object. Although the correct flow is among these candidates, it may not be the top-voted one ( 1st stage). c) We translate the source by those flow candidates and measure how well it aligns with the target by an additional voting ( 2nd stage). The flow receiving the highest number of votes is selected as the rigid flow, representing the underlying motion of the car.
Figure 2: Overview of RVLoss. a) A typical feedforward scene flow estimator Gθ takes as input a sequence of LiDAR scans ( ≥2 frames), P0−h,…,P0,P1 , and predicts per-point flows F0 . b) RVLoss employs a runoff vote mechanism to estimate rigid and free-form flows over clustered points during each forward pass, and uses the resulting flows as pseudo-labels. The runoff vote consists of a two-stage voting procedure built upon a shared operation, termed Guided Rigid Flow Estimation (GRF) . Specifically, GRF first uses the predicted flow F0 as guidance to select the top- k rigid flow candidates T0′ (Stage 1). Subsequently, GRF takes these top- k candidates as guidance and identifies the top- 1 rigid flow T0′′ (Stage 2). In addition to rigid flow, RVLoss also computes a free-form flow F0′′ as complementary supervision to relax the rigid-motion assumption and to mitigate the quantization error introduced by voting. The objective is to minimize the residual between the predicted flow and two supervision flows: the top- 1 rigid flow and the free-form flow.
Figure 3: Overview of GRF. a) Guided NN Search first translates the source points using the guidance flow and then finds their nearest neighbors in the target point cloud. The point-wise motion between each source point and its matched target point is used for the subsequent processing. b) Vote for Rigid Motion quantizes the point-wise flows within each cluster and casts them into a predefined voting space. The top- k flow candidates are then selected and assigned to all points within the cluster.
Backbone
Loss
#F input
#F loss
RT (s)
Dynamic Bucket-Normalized ↓
Three-way EPE (cm) ↓
Mean
CAR
O.V.
PED.
VRU
Mean
FD
FS
BS
Supervised methods
DeltaFlow
GT
5
-
8
0.103
0.083
0.132
0.122
0.075
1.76
3.16
1.31
0.81
SSF
GT
2
-
-
0.160
0.115
0.170
0.217
0.137
2.36
4.71
1.69
0.67
Self-supervised methods
DeFlow
SeFlow
2
2
7.2
0.313
0.236
0.295
0.441
0.273
4.76
10.93
2.23
1.12
Table 1: In-domain performance on the AV2 test split. Numerical results on scene flow prediction come from the AV2 2026 Scene Flow Challenge leaderboard. ‘#F input ’ denotes the number of input frames to the network. ‘#F loss ’ denotes the number of input frames during loss calculation. RT stands for runtime per sequence (around 157 frames) in seconds, retrieved from TeFlow [ 10 ] . ( Best , Second Best )
Figure 4: Cross-domain evaluation performance. We report the mean Dynamic Bucket-Normalized EPE on multiple datasets used in the AV2 2026 scene flow challenge [ 11 ] . DeltaFlow+RVLoss demonstrates consistent performance improvement over all baselines.
Figure 5: Qualitative comparison . Visualization of scene flow predictions from the DeltaFlow model trained with the TeFlow loss and the proposed RVLoss. Point colors denote flow direction, and color saturation indicates flow magnitude. Objects are expected to exhibit coherent color patterns. Compared with the TeFlow [ 10 ] loss, RVLoss produces more uniform coloring within each object (highlighted in boxes), demonstrating more coherent scene flow estimation.
Free-form Loss
Rigid Loss
Bidirect Optim
Dynamic Bucket-Normalized ↓
Three-way EPE (cm) ↓
Mean
CAR
O.V.
PD.
VRU
Mean
FD
FS
BS
✓
✓
✗
0.284
0.129
0.267
0.442
0.299
3.59
7.31
1.99
1.48
✗
✓
✓
0.338
0.138
0.268
0.604
0.346
3.70
8.27
1.69
1.14
✓
✗
✓
0.275
0.139
0.266
0.405
0.288
4.68
7.48
3.62
2.94
✓
✓
✓
0.262
0.121
0.246
0.405
0.277
3.47
6.84
2.05
1.53
Table 2: Ablation of each loss component on the Argoverse 2 val split [ 6 ] . ( Best , Second Best )
Appendix figures & tables3 assets
Supplementary material from the paper’s appendix.
Appendix
Method
Loss
nuScenes [ 53 ]
Waymo [ 24 ]
Mean
CAR
O.V.
PD.
VRU
Mean
CAR
PD.
VRU
DeFlow
SeFlow
0.544
0.396
0.653
0.726
0.419
0.351
0.212
0.551
0.289
VoteFlow
SeFlow
0.538
0.355
0.605
0.780
0.410
0.347
0.197
0.548
0.298
DeFlow ++
SeFlow ++
0.509
0.327
0.583
0.716
0.409
0.323
0.201
0.521
0.247
DeltaFlow
TeFlow
0.395
0.303
0.461
0.474
0.344
0.275
0.157
0.469
0.195
DeltaFlow
TeFlow (Re)
0.433
0.347
0.492
0.628
0.266
0.285
0.164
0.486
0.206
Appendix
Table 3: Dynamic Bucket-Normalized EPE ( ↓ ) on the nuScenes [ 53 ] val set and the Waymo [ 24 ] valid set. ( Best , Second Best )
Stage 1
Stage 2
Dynamic Bucket-Normalized ↓
Three-way EPE (cm) ↓
Mean
CAR
O.V.
PD.
VRU
Mean
FD
FS
BS
-
1
0.499
0.373
0.667
0.540
0.418
7.19
18.5
1.81
1.29
3
1
0.317
0.176
0.303
0.474
0.316
4.44
9.84
2.03
1.47
5
1
0.277
0.137
0.260
0.418
0.291
3.73
7.75
1.98
1.46
10
1
0.262
0.121
0.246
0.405
0.277
3.47
6.84
2.05
1.53
20
1
0.297
0.127
0.276
0.500
0.285
3.67
7.41
2.04
1.57
Appendix
Table 4: Ablation of k in voting modules on the Argoverse 2 val split [ 6 ] . ( Best , Second Best )
Method
Loss
Predicted flow ↓
(Refined) rigid flow ↓
(Refined) free-form flow ↓
Mean
CAR
O.V.
PD.
VRU
Mean
CAR
O.V.
PD.
VRU
Mean
CAR
O.V.
PD.
VRU
DeFlow
SeFlow
0.342
0.205
0.340
0.510
0.315
0.237
0.140
0.273
0.348
0.189
0.293
0.194
0.294
0.411
0.272
DeFlow ++
SeFlow ++
0.300
0.180
0.340
0.395
0.288
0.236
0.131
0.260
0.363
0.189
0.289
0.187
0.282
0.416
0.270
DeltaFlow
TeFlow
0.223
0.140
0.268
0.304
0.179
0.220
0.119
0.243
0.354
0.162
0.278
0.178
0.267
0.414
0.255
Appendix
Table 5: RVLoss as offline Flow Refinement. We use RVLoss to refine the predicted flows and thus obtain rigid flows and free-form flows. We compare these flows on the AV2 val split [ 6 ] using Dynamic Bucket-Normalized EPE. ( Best / Second Best in each row and category)
LiDAR scene flow is the task of estimating per-point 3D motion between consecutive point clouds. Recent methods achieve centimeter-level accuracy on popular autonomous vehicle (AV) datasets, but are typically only trained and evaluated on a single sensor. In this paper, we aim to learn general motion priors that transfer to diverse and unseen LiDAR sensors. However, prior work in LiDAR semantic segmentation and 3D object detection demonstrate that naively training on multiple datasets yields worse performance than single dataset models. Interestingly, we find that this conventional wisdom does not hold for motion estimation, and that state-of-the-art scene flow methods greatly benefit from cross-dataset training without architectural modification. We posit that low-level tasks such as motion estimation may be less sensitive to sensor configuration; indeed, our analysis shows that models trained on fast-moving objects (e.g., from highway datasets) perform well on fast-moving objects, even across different datasets. Informed by our analysis, we propose UniFlow, a feedforward model that unifies and trains on multiple large-scale LiDAR scene flow datasets with diverse sensor placements and point cloud densities. Our frustratingly simple solution establishes a new state-of-the-art on Waymo and nuScenes, improving over prior work by 5.1% and 35.2% respectively. Moreover, UniFlow achieves state-of-the-art accuracy on unseen datasets like TruckScenes and AEVAScenes, outperforming prior dataset-specific models by 30.1% and 22.5% respectively.
Siyi Li, Qingwen Zhang, Ishan Khatri +4
University of Pennsylvania · KTH Royal Institute of Technology · StackAV +1
LiDAR scene flow estimation is essential for autonomous driving, as it provides 3D motion for each point. Self-supervised approaches use static-dynamic classification to mitigate the imbalance between static and dynamic points, deriving targeted supervision. However, existing methods rely on sparse geometric observations for this classification, making them vulnerable to data sparsity and occlusions. The resulting noisy labels provide incorrect motion guidance and degrade scene flow learning. To address this, we introduce TrackCue, a tracking-guided framework for improving dynamic object representation in LiDAR scene flow estimation. In particular, TrackCue repurposes point tracking to obtain dense image-space trajectories anchored to LiDAR points, providing motion cues beyond sparse geometric observations. Furthermore, we present a visually consistent motion compensation strategy that compares the tracked trajectories with ego-induced rigid trajectories in the image plane, effectively isolating true object motion from ego-induced apparent motion. To transfer these isolated motion cues back to the LiDAR domain, we perform visual motion cue lifting, which associates ego-compensated image trajectories with LiDAR points for static-dynamic label refinement. As a result, TrackCue produces more accurate static-dynamic classification and provides more reliable supervision for scene flow learning. Experimental results show that TrackCue significantly improves the precision and F1 score of dynamic labels, leading to performance gains in self-supervised scene flow estimation.
LiDAR scene flow is the task of estimating point-wise motion from sequential point clouds. It is essential for autonomous driving, as it provides critical information about moving objects and their trajectories. In recent years, unsupervised methods have gained significant traction by eliminating the need for expensive point-wise annotations. However, despite incremental performance improvements, these methods uniformly rely on a single paradigm: training feed-forward networks through optimizing a loss function based on the Chamfer distance, a framework that inherently limits their progression. This paper presents CorrelationFlow, a new paradigm that reformulates LiDAR scene flow as a correlation maximization problem between Bird's-Eye View (BEV) images. Instead of using learnable parameters, our approach employs a chain of image processing operations to solve this problem. The learning-free nature of CorrelationFlow enables strong generalization across diverse domains. This capability is demonstrated by the second-place ranking in the unsupervised track of the Argoverse 2 Scene Flow Challenge at CVPR 2026, which featured a test set spanning five datasets that cover varied vehicle platforms and environments. Our results show that this classical, model-based approach can match or even surpass the performance of learning-based methods, thereby opening a new direction for unsupervised LiDAR scene flow research. The source code will be released.
Minh-Quan Dao, Yancong Lin, Shiming Wang +2
Nantes Université, LS2N · University of Nottingham · TU Delft +1