A Doppler radar measures one component of each reflecting point's velocity exactly: the part along the line from the sensor to the point. The sideways (tangential) part is invisible to it. Learned radar "scene flow" networks predict a 3D motion vector for every radar point between consecutive frames. Current networks treat this measured radial speed mostly as an input feature or as a soft training penalty. They then smooth the whole predicted vector across neighbouring points. That smoothing is what blurs the boundary between a moving cyclist and the static road beside it. We propose to stop smoothing what was measured.
The network's output is split by construction. The radial component of each point's flow is copied directly from that point's own Doppler reading, so it reaches the output unchanged, much as a skip connection carries fine detail around a compressed bottleneck. The network only predicts the tangential remainder. Neighbours share tangential information only when their Doppler readings are consistent with moving together. The result is a radar-only model at inference time. Its motion boundaries come from physics rather than from camera segmentation masks, and its learning capacity goes entirely to the part of motion the sensor cannot see.
Core Claim
Two failure modes look like separate problems in the file, and they share one cause. The first is tangential blindness: purely sideways motion produces zero Doppler shift. Hidden Gems: 4D Radar Scene Flow Learning Using Cross-Modal Supervision (CMFlow) names tangential supervision as an unresolved failure mode. The second is boundary bleeding: the spatial smoothness losses used by Self-Supervised Scene Flow Estimation With 4-D Automotive Radar (RaFlow) and Weakly Supervised Cross-Modal Learning for 4D Radar Scene Flow Estimation (IterFlow) push dynamic velocities into adjacent static clutter.
The shared cause is that existing networks treat the flow vector as a single unknown. Some of its information is observed per point, at full spatial resolution, and some is not. Our claim is that the measured part should be fixed at the output and the unmeasured part should be filled in by a learned, edge-aware propagation whose edges are themselves derived from Doppler. This should do two things. It should sharpen flow at static/dynamic boundaries and on small road users. It should also be at least no worse on tangential error than current methods, because the network no longer spends capacity re-predicting a quantity it was given.
Concretely, for point $i$ with unit line-of-sight direction $\hat r_i$, measured relative radial velocity $v_{r,i}$, and frame interval $\Delta t$, the predicted flow is
$$f_i = \Delta t\,(v_{r,i} + \delta_i)\,\hat r_i + (I - \hat r_i \hat r_i^{\top})\,t_i ,$$
where $t_i$ is the network's tangential prediction, which is projected so it cannot alter the radial part. The term $\delta_i$ is a small, bounded, learned correction that absorbs Doppler noise and, if needed, a discrete velocity-unwrapping offset. Because the radar reports velocity relative to the sensor, the radial part of the sensor-frame flow needs no ego-motion estimate at all.
Tangential information moves between neighbours $i$ and $j$ with weight
$$w_{ij} = \exp\!\big(-(v_{r,j} - \hat v_i \cdot \hat r_j)^2 / \sigma^2\big),$$
where $\hat v_i$ is the current full velocity estimate at $i$ and $\sigma$ is set from the sensor's Doppler resolution. In words, $i$ predicts what $j$'s Doppler reading should be if the two share a velocity. Message passing is cut wherever that prediction fails. A moving pedestrian next to a static kerb fails this test immediately, even with only two or three points, because their Doppler readings disagree after accounting for bearing. No camera mask is needed.
Relation to Prior Work
Prior work plays three roles here.
The closest work supplies a signal we reuse in a different way. RaFlow already exploits the per-point Doppler constraint, but as a radial displacement loss that competes with its smoothness and Chamfer losses during training. We move the constraint from the loss into the output parameterization, so it cannot be traded away. VISC: mmWave Radar Scene Flow Estimation using Pervasive Visual-Inertial Supervision and Full-Velocity Radar Returns by Radar-Camera Fusion both combine radial Doppler with 2D optical flow to recover full velocity. The second does so in closed form with a 3×3 inversion, but it needs a camera at inference and fails under occlusion and poor light. Our model uses no camera at inference. Cameras can still supervise the tangential part during training.
A second group motivates the gap. CMFlow reports the tangential failure mode. IterFlow restricts matching and smoothness to camera instance masks back-projected into 3D, which ties its boundaries to 2D segmentation quality and to good weather. TARS: Traffic-Aware Radar Scene Flow Estimation shows that instance rigidity breaks down when objects have only a handful of returns. It replaces rigidity with a grid-based traffic vector field, but that fixed grid limits sharpness at fine boundaries. Our Doppler affinity is a point-level, grid-free alternative to all three ways of deciding which points move together.
A third group supplies the architectural lesson and a constraint on it. Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections and The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation show that fine detail survives when it bypasses the bottleneck. Tiramisu's gains are largest on thin classes such as pedestrians and cyclists. The constraint is that in radar the high-frequency feature content is dominated by multipath ghosts, as RaFlow's density-weighted Chamfer loss and SDDiff: Boosting Radar Perception via Spatial-Doppler Diffusion both indicate. We therefore do not bypass raw features. We bypass only the one scalar per point that is a physical measurement.
Finally, Doppler-Aware LiDAR-RADAR Fusion for Weather-Robust 3D Detection reports that Doppler gets lost when simply concatenated with other channels, which supports keeping Doppler on a separate path. It also warns that automotive radars have a limited unambiguous velocity range, which is a direct risk for any method that trusts Doppler.
Research Program
We will build the decomposed output head and the Doppler-gated propagation on top of a standard point-based radar scene flow backbone. IterFlow's design is a reasonable template because it is lightweight: the file reports about 40× fewer parameters than CMFlow. Training will use the same weak or cross-modal supervision as the baseline it is compared to, so that only the architecture differs.
The propagation runs for a few iterations. Each iteration refreshes $\hat v_i$, recomputes $w_{ij}$, and aggregates tangential features over a ball neighbourhood. Static points can be handled with the same machinery: their Doppler readings agree with an ego-velocity fitted robustly to the scan, so the gate naturally groups them.
The bounded correction $\delta_i$ has two parts: a continuous term limited to a few Doppler resolution cells, and a classification over a small set of unwrapping offsets. Before relying on the hard pin, we will measure how often View-of-Delft (VoD) velocities approach the sensor's unambiguous limit. If aliasing is rare, the offset classifier is a safeguard. If it is common, unwrapping becomes a first-class sub-problem.
Analysis has three parts. First, we decompose error into its radial and tangential components. Second, we stratify error by the angle between an object's true velocity and its line of sight. Third, we measure error as a function of distance to the nearest static/dynamic boundary, using ground-truth boxes. These are the measurements that show where a gain comes from.
Baselines
The naive reference is a physics-only predictor that uses no learning. It fits ego-velocity to the scan's Doppler readings with RANSAC and assigns ego flow to inlier points. Outlier (moving) points receive only their radial flow $\Delta t\, v_{r,i}\hat r_i$, with zero tangential part. This baseline shows how much of the VoD metric is already explained by measured radial motion. Its tangential error defines the problem that remains. It will be run on the same frames and evaluated with the same metrics.
The strongest existing method on VoD in the file is TARS. TARS-superego reports an end-point error (EPE, the mean distance between predicted and true flow vectors) of 0.048 m, and TARS-ego reports EPE below 0.1 m, 0.038 m lower than CMFlow. TARS is the strongest competitor because it targets the same sparsity problem and claims reduced tangential underestimation. Its superego variant uses ego-motion compensation, so we will compare ego-compensated variants against each other and uncompensated variants against each other.
IterFlow is the strongest camera-boundary method. It reports a 34.7% EPE improvement and 13.6%/21.4% gains in the two accuracy metrics (AccS/AccR) over CMFlow. The file gives no absolute IterFlow EPE, so we will reproduce it under our protocol. CMFlow is the cross-modal reference.
All learned methods get the same training split, the same supervision sources, and the same hyperparameter search budget. Where code is not available we reimplement, and we report our reproduced numbers alongside the published ones. Beating TARS means overall EPE no worse than the matched TARS variant, together with a clear reduction in EPE on points within one metre of a moving/static boundary and on pedestrians and cyclists.
Three ablations isolate the key ingredient. The first uses the same backbone with an unconstrained 3D output plus a RaFlow-style radial loss, which tests output parameterization against a loss. The second keeps the decomposition but replaces the Doppler gate with a learned attention gate of equal capacity, which tests the physical edge signal against learned edges. The third keeps the decomposition and uses plain uniform smoothing.
Why not just…?
Why not just keep RaFlow's radial displacement loss and tune its weight? This is the most serious threat. A strong enough loss might achieve the same result as a hard pin. The argument for the pin is that the loss competes with smoothness and Chamfer terms precisely at boundaries, where neighbours disagree, so any finite weight lets some bleeding through. The first ablation settles this directly. We sweep the loss weight across the same tuning budget and compare boundary-band EPE. If a tuned loss matches the pin, the architectural claim is reduced to the gating contribution alone, and we will report it that way.
Why not just use standard U-Net-style skip connections or a learned attention gate? Skips carry raw early features, and in radar those include ghost reflections that RaFlow had to suppress with a dedicated density-weighted loss. A learned attention gate is a fairer competitor and is our second ablation. If it matches the Doppler gate, the physics adds nothing beyond inductive bias, which may still help most on small data. To check that, we also compare the two gates at reduced training-set sizes.
Why not just use camera instance masks, as IterFlow does, to stop bleeding? In clear daytime VoD scenes, this may work as well as our gate or better. Our method needs no camera at inference and does not inherit mask leakage. The evidence that camera-derived cues weaken in obscurants is indirect: it comes from VISC's smoke experiments, where CMFlow reached 0.322 m EPE against VISC's 0.275 m at 70% density. VoD has no adverse-weather flow labels. We will therefore also evaluate both methods with camera supervision restricted to frames whose detections are degraded, and report honestly if masks win in clear conditions.
Why not just apply smoothing less? Removing smoothness keeps boundaries sharp but leaves the tangential part almost unconstrained on objects with few points. The uniform-smoothing and no-smoothing variants of our ablations quantify this trade-off. The gated version is justified only if it beats both.
Proposed Experiments and Decisive Checks
The main benchmark is VoD, using its standard EPE, AccS, and AccR metrics. We add three diagnostics: radial versus tangential EPE, EPE stratified by motion-to-beam angle, and EPE within a one-metre band of moving/static boundaries. Where data access allows, we also evaluate on the proprietary-scale setting by running TARS's reported comparison protocol on any available high-resolution radar data. Otherwise we state that the evaluation is VoD-only.
Three outcomes would be decisive. First, boundary-band EPE and small-object EPE should drop relative to TARS and IterFlow, with no increase in tangential EPE on objects moving mostly sideways. If tangential EPE gets worse, the pin is starving the network of a useful signal. Second, parameterization should beat the tuned-loss ablation on boundary-band error. Third, the Doppler gate should beat equal-capacity learned attention, at least in the low-data regime.
A negative result on the first check together with a positive result on the third would still be informative. It would mean Doppler is a good grouping signal but a poor quantity to fix at the output.
References
- Hidden Gems: 4D Radar Scene Flow Learning Using Cross-Modal Supervision
- Self-Supervised Scene Flow Estimation With 4-D Automotive Radar
- Weakly Supervised Cross-Modal Learning for 4D Radar Scene Flow Estimation
- VISC: mmWave Radar Scene Flow Estimation using Pervasive Visual-Inertial Supervision
- Full-Velocity Radar Returns by Radar-Camera Fusion
- TARS: Traffic-Aware Radar Scene Flow Estimation
- Image Restoration Using Very Deep Convolutional Encoder-Decoder Networks with Symmetric Skip Connections
- The One Hundred Layers Tiramisu: Fully Convolutional DenseNets for Semantic Segmentation
- SDDiff: Boosting Radar Perception via Spatial-Doppler Diffusion
- Doppler-Aware LiDAR-RADAR Fusion for Weather-Robust 3D Detection