FIND helps LiDAR detectors adapt by separating stable shapes from changing sensor patterns

Why a familiar object can become unfamiliar

LiDAR represents a scene as scattered measurements in three-dimensional space. Object detectors turn those points into categories and bounding boxes. But changes in scanner configuration, location or weather can alter the measurements, weakening a detector trained elsewhere. Researchers call this mismatch a domain shift.

Dongxiao Li, Zixuan Hu and Ling-Yu Duan of Peking University propose FIND, short for Frequency Invariance Disentanglement. It adapts a pretrained detector while processing incoming, unlabeled scans. The supplied paper is an arXiv version-one preprint; its landing page reports acceptance at NeurIPS 2026. The results described here are the authors’ reported experiments, not an independent replication.

Frequency without sound

FIND’s frequencies describe geometric variation across a graph, not vibrations over time. The method takes points inside predicted object boxes, centers their coordinates and connects nearby points. A graph Fourier transform then represents the coordinates through patterns that vary more smoothly or more rapidly between connected points.

The authors argue that unwanted sensor effects cannot simply be removed by discarding high frequencies. Their analysis finds differences between datasets throughout the spectrum. Removing whole frequency ranges can also damage useful geometry: broad shape, connected structure or fine details.

To check whether those differences merely reflected different scenes, they also examined clean and artificially corrupted versions of the same KITTI scenes. They report changes across frequency bands even with underlying scene geometry held fixed. This supports their filtering rationale within the tested conditions, rather than proving a universal separation between geometry and interference.

Two views of the same scan

FIND uses two detector streams with learnable B-spline filters. B-splines are piecewise polynomial curves whose basis functions affect limited regions, giving the filters finer control over different frequency ranges than a single global response.

The invariant stream aims to identify stable geometric information. Its filtered representation determines which original points to retain, and its predictions supply pseudo-labels: automatically generated training targets rather than human annotations.

The specific stream extracts components intended to capture domain-dependent variation. It adds a randomly scaled version of those components to the original coordinates. Training encourages its predictions to agree with the invariant stream, making the detector less sensitive to those perturbations.

Two further constraints guide learning. One favors frequency bands whose spectral energy is relatively stable across incoming samples. The other encourages the two filters to sum to one, discouraging trivial solutions that discard everything.

That stability signal is a proxy, not direct knowledge of what an object means. Across three analyzed scenarios, the authors report strong correlations between within-domain spectral variability and cross-domain differences. FIND uses source-domain spectral statistics for initialization before deployment, but its online updates require neither source samples nor target labels.

What the benchmarks show

The main experiments use the SECOND-IoU detector. For cars at moderate detection difficulty, transferring from nuScenes to KITTI gives FIND a 3D average precision of 54.66, compared with 51.11 for MOS. Transferring from Waymo to KITTI gives 67.09 versus 64.16. Average precision summarizes detection performance; these differences are AP points, not percentages of objects correctly identified.

Across eight heavy corruption types in KITTI-C, FIND averages 59.62 AP, versus 57.48 for MOS. When combining a Waymo-trained detector with corrupted KITTI data, it averages 33.99 versus 32.62. FIND does not win every individual condition or category, and substantial gaps remain in difficult transfers.

A continual fog-to-rain-to-snow experiment, without resets or domain labels, produces an average only 0.37 AP points below separate adaptation to each corruption. That is evidence for this particular sequence, not unrestricted deployment robustness.

The practical limits

Adaptation adds computation. On one RTX 4090 with batches of eight, the authors report average processing time normalized per point cloud of 194 milliseconds, compared with 179 for CoTTA and 3,088 for MOS. This batched measurement does not establish single-frame latency or suitability for a vehicle’s timing requirements.

FIND also depends on predicted boxes to construct object graphs. Removing 10% of those boxes reduces its performance enough to fall slightly below MOS in the combined dataset-and-corruption benchmark. The authors acknowledge that large point-density changes may require manual filter tuning.

The contribution is therefore a benchmark-supported way to couple selective geometric filtering with online adaptation—not proof that stable object information can always be isolated, or that improved detection scores establish driving safety.