A new LiDAR registration method reported very high success when matching scans from different sensor types, including a 99.3% overall strict success rate on held-out HeLiPR pairs. The method, called CVSD-Reg, was also reported at 99.0% on nuScenes and 97.7% on KITTI under the study’s strict benchmark measure.
The results come from an arXiv version-1 preprint dated 20 Aug 2026. They show how this implementation performed on the reported datasets and test protocols; they do not establish that visual distillation causes the gains, or that the same performance would hold on other datasets, sensor designs or operating conditions.
A harder test for 3D scan matching
LiDAR registration means aligning two three-dimensional point clouds so that their relative position and orientation can be estimated. The study asked whether adding visual semantic information to LiDAR descriptors could make that alignment more robust to changes in point density, scan pattern, viewpoint and sensor characteristics.
CVSD-Reg uses a two-stage training framework. In the first stage, a frozen DINOv2 vision model transfers semantic information to a PTv3 LiDAR backbone using paired image-LiDAR data. The training includes hyperspherical alignment, which compares feature directions, and a rigid-view consistency objective that encourages matching representations across changed views.
The second stage removes the vision teacher and adapts the pretrained LiDAR backbone for registration using unposed LiDAR pairs. It learns correspondences between points and the relative pose while training sees different scan densities. At inference, the reported configuration uses no camera inputs, external pose solver or post-hoc ICP refinement.
That distinction matters to the interpretation of the result: images are used during pretraining, but the reported inference configuration is camera-free. The study therefore tests whether cross-modal training can leave a LiDAR-only system with a representation useful for scan matching.
Strong results across the reported benchmarks
The main evaluation covered 555 KITTI odometry pairs, 500 nuScenes validation pairs and 600 held-out HeLiPR cross-sensor pairs. The HeLiPR evaluation included 150 pairs per sensor. Additional diagnostic subsets were separate from those headline benchmark samples.
The primary measure was strict success: a registration counted as successful only within 0.5 metres of translation error and 1 degree of rotation error. The analysis also reported looser thresholds of 1 metre/2 degrees and 2 metres/5 degrees, along with median relative translation and rotation errors.
On HeLiPR, CVSD-Reg recorded reported strict success rates of 97.3% for Velodyne-16 and 100.0% for both Avia and Aeva, for 99.3% overall. BUFFER-X, a comparison method, was reported at 53.3%, 82.7% and 94.0% on those three sensors, with 82.5% overall.
The error summaries moved in the same direction on the HeLiPR comparison. Median relative translation error was reported as 0.107 metres for CVSD-Reg versus 0.269 metres for BUFFER-X, while median relative rotation error was 0.218 degrees versus 0.74 degrees.
The method was also reported at 97.7% strict success on KITTI, with median relative translation error of 0.047 metres and median relative rotation error of 0.153 degrees. On nuScenes, it reached 99.0%, compared with 97.0% for GeoTransformer and 89.6% for BUFFER-X.
What happened when parts of the system were removed
The clearest model-variant comparison came from a held-out HeLiPR ablation. A Stage-2-only control trained from scratch had 16.2% overall strict success, compared with 99.3% for the full CVSD-Reg system. The control’s sensor-specific rates were 58.0% on Ouster, 6.7% on Velodyne and 0.0% on both Avia and Aeva.
The comparison is consistent with an important role for the first training stage under this protocol, but it is not a clean causal test. The authors note that the Stage-2-only control’s random initialization created a harder optimization landscape that cannot be fully ruled out as an explanation for the gap.
A separate diagnostic comparison looked at projecting DINOv2 features at test time. On a fixed 100-pair KITTI subset, that approach had 24.0% strict success and required cameras, while CVSD-Reg had 98.0% and was camera-free. Because this was a diagnostic subset rather than the main 555-pair KITTI protocol, the figures should not be read as a replacement for the headline KITTI result.
The study also examined the ingredients of Stage 1. The complete objective, combining distillation, hyperspherical alignment, an information-contrastive term and rigid-view consistency, was reported as the best diagnostic configuration, 8.6 percentage points above distill-only.
In a dropout ablation using 50 LGR iterations, the stated setting of a 0.25 minimum keep ratio, equivalent to a maximum dropout strength of 0.75, was associated with Velodyne strict success changing from 91.3% to 97.3%. This was an augmentation comparison using the stated solver configuration, not one of the main benchmark protocols.
The evidence is promising but narrow
The headline pattern is limited to computational benchmark performance on the reported KITTI, nuScenes and HeLiPR protocols. The comparisons are between model variants and baselines rather than randomized experimental conditions, so the differences are associations tied to these particular data splits, sensor types and evaluation procedures.
The method still needs synchronized and calibrated image-LiDAR data during pretraining, even though it uses no camera at inference. Whether weaker or calibration-free cross-modal supervision could remove that requirement remains open.
The supplied analysis reports no uncertainty intervals, repeated-run variability, inferential tests, confidence intervals or p-values. It also reports no external validation on additional datasets or sensor architectures, so the benchmark percentages do not establish statistical significance, independent reproducibility or performance beyond the tested setup.
The diagnostic ablations used subsets or solver configurations that differed from the headline evaluations. Further evaluation would be needed to determine whether repeated-seed testing materially changes the reported benchmark and ablation estimates, and whether the pattern holds across more heterogeneous LiDAR datasets and sensor architectures.
For now, the study supports a narrower conclusion: this CVSD-Reg implementation was reported to register the tested LiDAR pairs successfully across several within-sensor and cross-sensor protocols, while its broader value for mapping, localization, autonomy or other downstream outcomes was not tested in the supplied analysis.
Paper data and sources
Original title: CVSD-Reg: Cross-Modal Visual Semantic Prior Distillation for Robust LiDAR Registration
Authors: Eunsoo Im, Junghun Suh, Gyeonggwan Lee, Seunghwan Hong
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-20
DOI: Not available
Original paper · Full text