NeurIPS 2026

RVLoss Runoff Vote Loss for Self‑Supervised
LiDAR Scene Flow Estimation

Shiming Wang1· Liangliang Nan1· Julian F. P. Kooij1· Holger Caesar1· Yancong Lin2,*

1TU Delft, the Netherlands 2University of Nottingham, UK *Corresponding author

This is a preliminary project page. Code, the talk video and the poster will be added soon.

Overview

Runoff vote illustration: nearest-neighbor matching yields several dominant flows within a car; translating the source by each candidate and voting again recovers the correct rigid flow.
The runoff vote. (a) A rigidly moving car has one consistent flow. (b) Nearest-neighbor matching, as used in Chamfer losses, breaks rigidity and produces several dominant flow candidates; the correct one may not win the first vote. (c) RVLoss translates the source by each candidate, re-matches it to the target and votes again. The candidate with the most support becomes the cluster's rigid flow.

Abstract

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 Scene Flow 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% and reducing the mean dynamic normalized end-point error from 0.202 to 0.163. Moreover, cross-dataset evaluations demonstrate consistent performance improvements across four additional datasets.

Highlights

0.163 vs 0.202
Mean dynamic normalized EPE on the AV2 test split
Same DeltaFlow backbone, supervised by RVLoss instead of the TeFlow loss. Lower is better.
5 / 5
Datasets with the lowest error
Among the compared self-supervised models on AV2, Waymo, nuScenes, TruckScenes and AevaScenes, all trained on AV2 only.
0 ms
Extra inference cost
The runoff vote runs only during training. The scene flow backbone and its inference are unchanged.

Qualitative results

TeFlow lossscene flow
Ground truthscene flow
RVLoss (ours)scene flow
TeFlow losserror
Closer to GTRVLoss vs. TeFlow loss
RVLoss (ours)error
RVLoss closer to ground truth TeFlow loss closer Each bar is one frame: mean end-point error of the TeFlow loss model minus that of RVLoss, over moving points (ground truth ≥ 0.1 m per frame). Click or drag on the bars to choose a frame; arrow keys step one frame.

Scene flow

Flow color wheel
Hue is the direction of motion (right = driving direction), saturation the speed, full at 2 m per frame. Gray points are not moving (< 0.1 m per frame) according to that model.

Error per point

00.250.50.75≥ 1 m
End-point error against the ground truth, in meters per frame. Below 0.05 m is drawn gray; darker gray points have no comparable ground truth.

Closer to ground truth

RVLoss is closer
TeFlow loss is closer
within 5 cm of each other
Color strength grows with the gap, full at 0.5 m.

Argoverse 2 validation set. Both models use the DeltaFlow backbone; one is trained with the TeFlow loss, the other with RVLoss. Bird’s-eye view of 60 × 60 m around the ego vehicle (center, driving to the right), ground points removed, played at the original 10 Hz. Highlight clips are 5-second windows selected where RVLoss’s advantage is largest; full sequences are unedited and include frames where the TeFlow loss model is closer to the ground truth.

BibTeX

@inproceedings{wang2026rvloss,
  title     = {{RVLoss}: Runoff Vote Loss for Self-Supervised {LiDAR} Scene Flow Estimation},
  author    = {Wang, Shiming and Nan, Liangliang and Kooij, Julian F. P. and Caesar, Holger and Lin, Yancong},
  booktitle = {Advances in Neural Information Processing Systems (NeurIPS)},
  year      = {2026}
}