CorrelationFlow: A Training-Free Geometric Approach for LiDAR Scene Flow Estimation

The gist The CorrelationFlow framework presents a training-free geometric approach to LiDAR scene flow estimation by formulating it as correlation maximization between bird’s-eye-view occupancy images, demonstrating that classical computer vision operations can solve a substantial share of the scene flow problem <ref:2607.29237#pg2>.

Core Methodology

CorrelationFlow reduces scene flow estimation to two textbook operations: connected-component labeling and correlation maximization on bird’s-eye-view occupancy images. Objects are isolated as spatio-temporal connected components, their motions recovered as correlation peaks, and the resulting velocities propagated to all member points. This approach is based on the intuition that while an object's position changes over a short time period, its point cloud appearance remains relatively constant.

Problem Formulation and Scene Flow Estimation

The problem defines scene flow as the displacement of a point between two consecutive time steps, expressed in the body frame of the ego vehicle at time (t + 1). To estimate this flow for an object, the method orthogonally projects point clouds onto 2D planes, such as the Bird’s-Eye-View (BEV) and a sectional view. The horizontal components are estimated by finding the 2D translation ∆ that maximizes the overlap between its foreground and that of I O t+1. This translation is recovered via normalized cross-correlation (NCC) applied to binary images derived from the point clouds.

Key Components for Robustness

The framework incorporates several components to address inherent limitations of geometric methods. The spatio-temporal clustering uses connected components on the BEV projection, which is enhanced by morphological dilation with a 3×3 structuring element of ones, 13×3, to I O t C 1 C 2 in C 2 only under 8-connectivity. This addresses occlusion and sparsity by bridging gaps between disjoint parts of an object. Furthermore, a coarse-to-fine strategy is employed for fast-moving objects, applying the clustering at a large pixel size s and repeating at progressively finer resolutions.

Variants and Performance Analysis

The paper introduces two main variants. The dense variant uses a temporal window to recover per-component motion by correlation maximization, robustified by a median over multiple baselines. The key-point variant, CorrelationFlow-KeyPoints, operates on a single sweep pair by matching boundary key points derived from the distance transform and filtering matches using Lowe’s ratio test. This variant shows that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps.

Experimental Validation

The method was evaluated on the Argoverse 2 2026 Scene Flow Challenge, where CorrelationFlow ranked second among unsupervised methods. Ablation studies show that increasing the number of past point clouds (h) reduces error for all classes, with the largest reduction occurring between h = 0 and h = 3. The comparison against state-of-the-art unsupervised methods confirms that classical geometric operations remain competitive with learned methods across heterogeneous sensors and platforms.

Limitations

The limitations include the fact that the BEV projection in the base version discards vertical motion and conflates points at different heights within a pixel. Additionally, both variants estimate translation only, meaning they do not capture intra-frame rotation or articulation. The performance is also tied to assumptions regarding piecewise-rigidity, where undersegmentation or oversegmentation can degrade the flow estimation. Finally, the method presumes accurate ego-motion compensation; pose errors translate directly into spurious flow.

The final results indicate that neither variant requires training data, labels, or learned parameters and both expose a single interpretable correlation operation at their core. The CorrelationFlow-Keypoint variant achieves a mean Dynamic Bucket-Normalized EPE of 0.2845, improving over the dense CorrelationFlow at h = 0 (0.3524) by 19.3%. This suggests that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning. The range-robustness of CorrelationFlow-Keypoints degrades less than the dense variant at long range. This confirms that matching lightweight boundary descriptors on a single sweep pair recovers most of the accuracy that the dense variant obtains only after aggregating several past sweeps. This transparent baseline is valuable both as a practical, reproducible estimator and as a yardstick for how much of the problem is solvable without learning.

1. textbfKeypoint-based flow estimation for single-sweep pairs (CorrelationFlow-KeyPoints): This variant operates on a single sweep pair and thus requires no temporal window, allowing for real-time or low-latency scene flow estimation without needing to store or process long historical sequences, which is critical for autonomous vehicle applications.

2. textbfRobust spatio-temporal clustering via connected components: The method replaces density-based clustering with connected components on BEV occupancy maps and includes a validation step where if the longer side of the bounding box of a connected component C i exceeds a threshold d max, reflecting the maximum expected object size, the component is eroded, which mitigates errors from occlusion and sparsity.

3. textbfAdaptive multi-scale strategy for fast-moving objects: The coarse-to-fine approach allows the system to handle fast motion by applying Alg. 3 at a coarse resolution, i.e., with a large pixel size s and repeating the process at finer resolutions until convergence, specifically addressing objects that can travel farther than their own footprint between time steps.

4. textbfIncreased robustness to sensor heterogeneity: The method is demonstrated to generalize across five heterogeneous datasets (AEVA, Argoverse 2, nuScenes, TruckScenes, and Waymo Open Dataset), with CorrelationFlow ranking second among unsupervised methods on the Argoverse 2 benchmark at CVPR 2026. This confirms that data-centric approaches are not the only path forward.

5. textbf Enhanced error reduction via past point clouds: Utilizing past point clouds (varying h from 0 to 8) shows that aggregating more data makes small and slow-moving objects, such as pedestrians, more visible in the BEV, leading to a significant reduction in mean error for all classes, with the largest gains occurring between h = 0 and h = 3.

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.