A new preprint reports that Ex-Sim(3)-Reg, a proposed method for pruning image-to-point-cloud correspondences, reported higher scores than a baseline in a disturbed benchmark. On 7-Scenes, its mean inlier ratio was 0.548 against 0.266 for the baseline, while mean registration recall was 0.472 against 0.292. The paper labels those differences as gains of 28.2% and 18.0%, respectively.
The result was not confined to 7-Scenes. In disturbed TUM tests, mean IR was 0.485 for Ex-Sim(3)-Reg versus 0.230 for the baseline, and mean RR was 0.302 versus 0.117. The reported gains were 25.5% for IR and 18.5% for RR.
How the pruning works
Ex-Sim(3)-Reg is described as a fast, non-learning pruning method. It approximates an extended version of Sim(3) registration, an alignment model built around rotation, translation and scale, using scale-parameterized compatibility graphs and a theoretical analysis. The approach searches for compatible candidate correspondences before estimating a pose, meaning the position and orientation of the aligned view.
The extended objective gives each correspondence its own scale value while also choosing an inlier indicator, rotation and translation. It minimizes a weighted squared distance between a reference 3D point and a depth-lifted 3D point after transformation. Put simply, the optimization can adjust scale separately for each candidate match.
The pipeline has three stages. It samples possible scale hypotheses, searches for candidate inliers and estimates the pose, then refines that estimate with RANSAC-based P3P. The authors present the workflow as a fast, non-learning route from candidate correspondences to a registration result.
The test moved beyond one dataset
The benchmark generated correspondence sets with a pre-trained MATR model under disturbed, unseen and low-inlier conditions, then compared methods across 7-Scenes, RGBD-v2, ScanNet and TUM. In a low-inlier setting, fewer candidate correspondences are correct. The main measures were inlier ratio, or the share of correct inliers among candidate correspondences, and registration recall, which evaluates whether the alignment meets a distance criterion. By default, the IR threshold was 10 cm in 3D and the RR threshold was 5 cm based on average 3D distance.
On a 7-Scenes to RGBD-v2 generalization test, mean IR was 0.455 with Ex-Sim(3)-Reg compared with 0.250 for the baseline. Mean RR was 0.383 versus 0.252. The paper reports gains of 20.5% and 13.1%, respectively.
On 7-Scenes to ScanNet, mean IR was 0.662 versus 0.409 for the baseline, and mean RR was 0.635 versus 0.431. The annotated gains were 25.3% for IR and 20.4% for RR.
Under low-inlier RGBD-v2 conditions, mean IR was 0.327 versus 0.166 and mean RR was 0.542 versus 0.295. The reported gains were 16.1% for IR and 24.7% for RR. The authors' theoretical claim is narrower than a promise to recover every correct match: it applies to a candidate subset of the true inliers.
The pruning step was also added to five listed image-to-point-cloud registration methods in a cross-domain ScanNet comparison. RR rose from 0.842 to 0.943 for MATR, from 0.780 to 0.932 for FreeReg, from 0.901 to 0.950 for Top-I2P, from 0.898 to 0.957 for MinCD, and from 0.914 to 0.963 for Diff-I2P. IR increased in every listed pair as well.
Accuracy came with a cost
An ablation of the number of scale sub-intervals showed an accuracy-runtime trade-off. At six sub-intervals, IR/RR were 0.542 and 0.394, with a runtime of 65.4 ms. At eight, IR/RR were 0.548 and 0.472, while runtime was 73.6 ms. The text says performance saturates when the number of sub-intervals is greater than six and says it adopts six.
On noisy TUM, an additional pose-error comparison gave the proposed method a relative translation error of 0.036 m and a relative rotation error of 1.824 degrees, lower than all listed comparators. The first measure captures translational mismatch and the second rotational mismatch, so both are error measures in which lower values are better within that comparison.
The document is an arXiv preprint, version 1 dated 28 Aug 2026. Its theoretical recovery result is explicitly limited to a candidate subset of the true inliers, and the authors identify refinement of the extended compatibility graphs as future work. The paper therefore does not claim that all true inliers have been recovered.
Paper data and sources
Original title: Ex-Sim(3)-Reg: 2D-3D Correspondence Pruning via Extended Sim(3) Registration
Authors: Pei An, Muyao Peng, Junfeng Ding et al.
Journal/Repository: arXiv
Status: Preprint, not yet peer-reviewed
First online: 2026-08-28
DOI: Not available
Original paper · Full text