TeFlow: Self-Supervised Multi-Frame LiDAR Scene Flow

• The paper demonstrates that multi-frame candidate pooling with temporal ensembling narrows the accuracy gap to optimization-based methods while maintaining real-time performance.
• TeFlow employs a consensus voting mechanism that combines directional consistency and reliability weighting to filter out noisy motion cues from occlusions and sparse sampling.
• Empirical results on datasets like Argoverse 2 and nuScenes show significant speedups and substantial improvements in dynamic object motion prediction.

TeFlow is a self-supervised, feed-forward scene flow method for LiDAR point clouds that is designed to make multi-frame supervision work for real-time models. Its central premise is that two-frame supervision derived from point correspondences is unreliable and often breaks down under occlusions, sparse sampling, sensor noise, and ambiguous motion, whereas multi-frame supervision can provide more stable guidance only if temporally inconsistent motion cues are filtered rather than accumulated naively. TeFlow therefore mines temporally consistent supervision across multiple frames and uses that supervision to train a feed-forward model that remains real-time at inference while narrowing the accuracy gap to optimization-based scene flow systems (Zhang et al., 22 Feb 2026).

1. Problem setting and methodological scope

Scene flow estimation predicts the $3$-D motion of each point between frames. In the setting targeted by TeFlow, the main difficulty is not runtime at inference but the quality of the self-supervisory signal used during training. Existing self-supervised feed-forward methods are fast, but they are usually trained from only two consecutive frames, and that supervision is fragile whenever correspondences disappear, change abruptly, or become unreliable because of occlusion or sparse LiDAR sampling (Zhang et al., 22 Feb 2026).

TeFlow is motivated by the observation that simply extending a two-frame objective to more frames does not solve the problem. The stated reason is that point correspondences vary abruptly across frames, producing inconsistent signals. Naive multi-frame objectives therefore accumulate contradiction as well as information. The method is explicitly framed as a bridge between two regimes: optimization-based methods, which can use many frames and achieve excellent accuracy but are too slow for real-time deployment, and feed-forward methods, which are efficient but under-supervised (Zhang et al., 22 Feb 2026).

The method assumes that a scene has been partitioned into static points and dynamic clusters $\mathcal{C}_j$. Static points are supervised to have near-zero residual flow, while each dynamic cluster is assumed to move rigidly. This cluster-level formulation is the basis on which temporally consistent motion hypotheses are collected, scored, and aggregated.

2. Temporal ensembling and candidate-pool construction

The core mechanism in TeFlow is temporal ensembling of motion cues for each dynamic cluster. Rather than trusting a single pairwise correspondence or a single neighboring frame, the method constructs a candidate pool of motion hypotheses and then mines the most temporally consistent subset (Zhang et al., 22 Feb 2026).

For a dynamic cluster $\mathcal{C}_j$, the first candidate is an internal estimate derived from the model’s current prediction: $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$ This internal candidate acts as an anchor in the sense described in the paper: it keeps supervision grounded in the model’s current state.

External candidates are then extracted from neighboring frames $t'\in\{t-h,\dots,t-1,t+1\}$. For each point $\mathbf{p}_k$ in the cluster, TeFlow finds its nearest neighbor in the dynamic points of frame $t'$ and converts the displacement into a per-frame velocity: $$\mathbf{f}^{t'}_{\mathcal C_j,k} = \frac{\mathcal{NN}(\mathbf{p}_k, \mathcal{P}_{t',d}) - \mathbf{p}_k}{t' - t}.$$ Here $\mathcal{NN}(\cdot)$ denotes nearest-neighbor search and $\mathcal{P}_{t',d}$ is the set of dynamic points in frame $\mathcal{C}_j$0.

The full candidate pool is

$\mathcal{C}_j$1

The paper states that this pool contains $\mathcal{C}_j$2 candidates. The construction is explicitly designed so that some candidates may be corrupted by occlusion or noise without dominating the supervisory target.

A common misconception addressed by the method is that “more frames” are automatically beneficial. TeFlow rejects that interpretation. Its claim is narrower and more technical: more frames help only when temporally inconsistent cues are filtered by a consensus mechanism rather than injected directly into the loss.

3. Consensus voting, reliability weighting, and final supervision

Once the candidate pool has been built, TeFlow computes a consensus score by combining directional consistency and candidate reliability (Zhang et al., 22 Feb 2026). Directional consistency is encoded by a binary agreement matrix $\mathcal{C}_j$3: $\mathcal{C}_j$4 with $\mathcal{C}_j$5, corresponding to a $\mathcal{C}_j$6 angular threshold.

Reliability is encoded by

$\mathcal{C}_j$7

where $\mathcal{C}_j$8, $\mathcal{C}_j$9 is the temporal offset of the candidate, and the magnitude term gives more influence to larger displacements.

The consensus score is then

$\mathcal{C}_j$0

and the winner is selected by

$\mathcal{C}_j$1

TeFlow does not use the winner alone. It aggregates all candidates that agree with the winner: $\mathcal{C}_j$2 This $\mathcal{C}_j$3 becomes the supervisory target for cluster $\mathcal{C}_j$4.

The distinction between selection and aggregation is important. The winner identifies the most supported motion hypothesis, while the final average suppresses outliers and reduces sensitivity to noisy nearest-neighbor matches. This suggests that TeFlow’s stability derives not from any single cue source, but from agreement structure across the temporal candidate set.

4. Loss design, backbone context, and implementation setting

TeFlow’s principal learning signal for dynamic objects is the dynamic cluster loss $\mathcal{C}_j$5, which contains both a point-level term and a cluster-level term: $\mathcal{C}_j$6 The rationale given in the paper is that point-level averaging alone is dominated by large objects, whereas cluster-level averaging gives smaller objects, including pedestrians and other VRUs, fair influence.

The total objective is

$\mathcal{C}_j$7

where $\mathcal{C}_j$8 penalizes non-zero residual flow on static background points and $\mathcal{C}_j$9 provides geometric consistency through multi-frame Chamfer-style alignment (Zhang et al., 22 Feb 2026).

In architectural terms, TeFlow is built on the multi-frame DeltaFlow backbone, and the paper also applies it to Flow4D to argue that the supervision mechanism is backbone-agnostic. Static/dynamic segmentation is obtained from DUFOMap, and dynamic clusters are precomputed with HDBSCAN. The voxel grid is $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$0 over a $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$1 m region at $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$2 m resolution. Training uses Adam for $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$3 epochs, with learning rate $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$4, batch size $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$5 across $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$6 NVIDIA RTX 3080 GPUs, and a training time of about $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$7–$$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$8 hours per dataset. The best temporal window is five frames.

5. Empirical performance and ablation findings

The reported evaluations cover Argoverse 2, nuScenes, and Waymo, with the strongest quantitative claims given for Argoverse 2 and nuScenes (Zhang et al., 22 Feb 2026). On the Argoverse 2 test set, TeFlow achieves Three-way EPE: 3.57 cm, matching Floxels at 3.57 cm while running at 8 s per sequence versus 24 min for Floxels, which the paper summarizes as a 150× speedup. It also reports that Dynamic Bucket-Normalized EPE improves by 22.3% over SeFlow++, and pedestrian error is reduced by 31%.

On the nuScenes validation set, the paper reports Three-way EPE: 4.64 cm and the best dynamic bucket-normalized score of 0.395. It further states improvements of 22.4% over SeFlow++ and a pedestrian error reduction of 33.8%. These results are used to support the paper’s main claim that TeFlow performs on par with leading optimization-based methods while retaining feed-forward efficiency.

The ablations are tightly aligned with the method’s design logic. Performance peaks at 5 frames, while adding more distant frames can degrade results because those frames introduce noise. Using only the internal candidate is the weakest variant; using only external candidates is better; and combining both is best. Among the voting components, removing directional consistency hurts the most, removing reliability weights also hurts, and removing aggregation degrades performance further. The full loss $$\hat{\mathbf{f}}_{\mathcal C_j} = \frac{1}{|\mathcal{C}_j|} \sum_{\mathbf{p}_i \in \mathcal{C}_j} \hat{\mathbf{f}}_i .$$9 performs best.

These ablations do not merely compare hyperparameters; they directly test the article’s central methodological claim. In the paper’s own terms, TeFlow’s contribution is not “multi-frame input” in general, but temporally consistent supervision extracted from a multi-frame candidate pool.

6. Nomenclature, disambiguation, and related usages

The name TeFlow is context-dependent in the arXiv record and is not unique to scene flow estimation. In the 2026 paper titled “TeFlow: Enabling Multi-frame Supervision for Self-Supervised Feed-forward Scene Flow Estimation”, it is the exact name of a LiDAR scene flow method (Zhang et al., 22 Feb 2026). In a different context, “Renormalization group for open quantum systems using environment temperature as flow parameter” uses TeFlow / T-flow to denote a real-time renormalization-group scheme in which reservoir temperature is the flow parameter (Nestmann et al., 2021). Another separate usage appears in “Transport-Embedded Neural Architecture: Redefining the Landscape of physics aware neural models in fluid mechanics”, where TeFlow refers to a transport-embedded neural architecture also called TENN (Jafari, 2024).

This suggests that the label should not be read as a stable cross-domain acronym. In current usage, the most specific and unambiguous meaning of TeFlow is the 2026 scene flow method for mining temporally consistent multi-frame supervision in self-supervised feed-forward LiDAR scene flow estimation (Zhang et al., 22 Feb 2026).