ScanNet-GSReg: Gaussian Splatting Registration Benchmark

Updated 7 July 2026

ScanNet-GSReg is a benchmark that evaluates alignment between independently trained Gaussian Splatting scene representations derived from indoor ScanNet data.
It employs a canonical coarse-to-fine pipeline that combines point-cloud-style alignment with image-guided refinement, measuring errors in rotation, translation, and scale.
The dataset comprises 1,379 indoor scene pairs created from overlapping image subsequences, addressing challenges like partial overlap and coordinate-frame inconsistency.

Searching arXiv for the core paper and closely related work on ScanNet-GSReg. ScanNet-GSReg is a scene-level registration benchmark derived from ScanNet for evaluating alignment between independently trained Gaussian Splatting (GS) scene representations rather than raw scans or conventional point clouds. It was introduced together with GaussReg, which formulates registration between two GS reconstructions of overlapping indoor scene parts as estimation of a similarity transform and couples coarse point-cloud-style alignment with fine image-guided refinement (Chang et al., 2024). The benchmark consists of 1,379 scene pairs built from ScanNet indoor scenes, with each pair containing two separately optimized GS models reconstructed from overlapping image subsequences whose camera extrinsics have been independently perturbed to simulate inconsistent world coordinate frames.

1. Origins and rationale

ScanNet-GSReg was motivated by a specific gap in 3D registration research: while point cloud registration had reached a nearly mature stage under deep learning, registration of radiance-field-style scene representations had received little exploration, largely because geometric relationships between implicit scene models are difficult to model directly (Chang et al., 2024). Existing approaches typically converted implicit representations into explicit ones before registration. Gaussian Splatting altered this landscape by representing scenes with explicit 3D Gaussians, thereby making direct scene-level registration a concrete benchmarkable problem.

The benchmark is designed around overlapping reconstructions rather than single-scene reconstruction quality alone. ScanNet already provides RGB-D sequences together with camera intrinsics and extrinsics, but ScanNet-GSReg converts these sequences into pairs of independently trained GS models that must be aligned. This setup makes the evaluation target explicit: not merely whether a GS model renders well, but whether two GS sub-scenes can be brought into consistent correspondence under scale, rotation, and translation.

Relative to prior registration datasets such as 3DMatch and ETH, ScanNet-GSReg changes both the representation and the difficulty profile. It is built around explicit GS reconstructions rather than raw scans, and the GS-derived point clouds used in coarse registration are described as noisier and only rough geometric approximations. The benchmark also explicitly evaluates scale through Relative Scale Error, reflecting monocular scale uncertainty; this differs from many prior registration settings that assume fixed scale.

2. Dataset construction and composition

ScanNet-GSReg is built from the ScanNet corpus by sampling two continuous image subsequences per scene, training a separate GS model on each subsequence, and recording the known relative transform induced by independent perturbations of their camera extrinsics (Chang et al., 2024). The resulting benchmark contains 1,379 pairs, partitioned into 1,297 training pairs and 82 test pairs.

Split	Pairs
Train	1,297
Test	82

These totals arise from starting with 1,513 ScanNet training scenes and 100 ScanNet test scenes, then excluding cases with failed initial point-cloud generation or unsuccessful GS reconstruction. From each retained scene, two continuous subsequences are randomly sampled. Each subsequence has length 80–120 images, the frame sampling interval is 1–5 frames, and the overlap ratio between the two sequences is 0.2–0.8, defined as the proportion of repeated images across the two sequences.

The creation pipeline consists of four steps. First, two continuous subsequences are sampled. Second, random transformations are applied independently to the two subsequences’ camera extrinsics so that the resulting reconstructions inhabit different world coordinate frames; the relative transform between these perturbations becomes the ground truth registration target. Third, two GS models are trained separately from the images and camera parameters of the two subsequences. Fourth, each GS model is optimized for 10,000 iterations. The benchmark therefore embeds controlled inconsistency in world coordinates while preserving exact ground-truth similarity transforms.

The scenes are indoor because the source corpus is ScanNet. Diversity is introduced through substantial variation in overlap ratio and sampling interval, as well as through simulated differences in scale, rotation, and translation caused by the independent world-frame perturbations. A plausible implication is that ScanNet-GSReg evaluates not only overlap robustness but also sensitivity to partial coverage and coordinate-frame inconsistency.

3. Representation, annotations, and task definition

Each scene in ScanNet-GSReg is represented as a GS model whose constituent 3D Gaussians store position $(x,y,z)$ , opacity $\alpha$ , rotation, scale, and spherical harmonics coefficients (Chang et al., 2024). For the coarse registration stage, colors $(r,g,b)$ are computed via the spherical harmonic functions associated with each Gaussian. Gaussians are then filtered by opacity, and only those with $\alpha > 0.7$ are retained to form the point-cloud-like input. The input channels per point to the coarse pipeline are $(x,y,z,\alpha,r,g,b)$ .

Camera supervision is preserved at the sequence level. For sequence $A$ , training camera poses are retained as $C_i^A=(R_i^A,T_i^A)$ ; for sequence $B$ , they are retained as $C_j^B=(R_j^B,T_j^B)$ . ScanNet camera intrinsics are also available and are used both for GS training and for later rendering during fine registration. By contrast, the paper does not specify the dataset’s file formats, directory layout, or renderer internals such as per-ray accumulation equations.

The registration target is a similarity transform aligning scene $B$ to scene $\alpha$ 0, parameterized as $\alpha$ 1, where $\alpha$ 2, $\alpha$ 3, and $\alpha$ 4. In matrix form, the rigid component may be written as $\alpha$ 5, and inclusion of $\alpha$ 6 yields a $\alpha$ 7 transform. The paper also gives the fused-coordinate update

$\alpha$ 8

together with

$\alpha$ 9

Ground truth is defined by the recorded relative transform between the two independently applied random extrinsic transformations during pair creation, so the benchmark directly supervises scale, rotation, and translation.

4. Canonical coarse-to-fine workflow

The benchmark was introduced with a canonical coarse-to-fine registration pipeline, GaussReg, that uses both GS-derived geometry and GS-rendered imagery (Chang et al., 2024). In the coarse stage, the filtered GS points are processed by a network following GeoTransformer’s pipeline: KPConv-FPN is used for multi-scale feature extraction, superpoint matching is performed, and ICP computes the coarse transform $(r,g,b)$ 0. Training uses the overlap-aware circle loss and the point matching loss from GeoTransformer, together with rotation, translation, and scaling augmentation to address monocular scale uncertainty. The exact loss formulas are not restated in the paper.

The fine stage is image-guided and begins with overlap image selection. Thirty training camera poses are uniformly sampled from each GS sequence. The coarse transform is applied to the cameras of scene $(r,g,b)$ 1. Candidate overlap views are then selected in three steps: first, the cosine of the angle between candidate camera orientations is computed and the top- $(r,g,b)$ 2 closest pairs are retained with $(r,g,b)$ 3; second, low-resolution depth maps are rendered from both GS models and mutual visibility is estimated, after which the closest pair is selected by maximizing averaged mutual visibility; third, neighborhoods around that pair are chosen and images $(r,g,b)$ 4 and $(r,g,b)$ 5 are rendered.

Fine feature extraction uses an image-guided 3D feature representation. A cost volume is constructed via differentiable homography from reference and source features under $(r,g,b)$ 6 depth hypotheses $(r,g,b)$ 7. The minimum and maximum distances for the cost volume are computed automatically from a rendered depth map of the reference image. After 3D CNN regularization, a probability volume $(r,g,b)$ 8 and a feature volume $(r,g,b)$ 9 are produced. For a pixel $\alpha > 0.7$ 0, the selected depth index is

$\alpha > 0.7$ 1

The corresponding feature, depth, and confidence are then computed as

$\alpha > 0.7$ 2

$\alpha > 0.7$ 3

$\alpha > 0.7$ 4

High- and low-resolution confidence-filtered features are formed as

$\alpha > 0.7$ 5

$\alpha > 0.7$ 6

These features are projected into the 3D coordinate frame of $\alpha > 0.7$ 7 using the corresponding depth map, and matched to refine the registration to $\alpha > 0.7$ 8.

For the fine stage, the depth loss is

$\alpha > 0.7$ 9

The registration loss is the same as in the coarse stage. A total loss $(x,y,z,\alpha,r,g,b)$ 0 with $(x,y,z,\alpha,r,g,b)$ 1 is referenced, but the explicit formula is not shown. In the reported experiments, $(x,y,z,\alpha,r,g,b)$ 2 depth hypotheses are used, and $(x,y,z,\alpha,r,g,b)$ 3 images per GS are rendered for the fine stage. Coarse and fine networks are trained separately for 40 epochs with batch size 1 and learning rate $(x,y,z,\alpha,r,g,b)$ 4 decayed exponentially by 0.05 per epoch, and the coarse input is limited to 30k points.

5. Evaluation protocol and benchmark results

ScanNet-GSReg evaluates similarity-transform registration using Relative Rotation Error, Relative Translation Error, Relative Scale Error, Success Ratio, and runtime (Chang et al., 2024). The reported definitions are

$(x,y,z,\alpha,r,g,b)$ 5

$(x,y,z,\alpha,r,g,b)$ 6

$(x,y,z,\alpha,r,g,b)$ 7

Success Ratio is the fraction of pairs considered successfully registered, but the exact success criterion is not specified. Reported runtime includes the preprocessing time needed to obtain the point cloud from GS.

The benchmark results reported for ScanNet-GSReg compare image-based, point-cloud-based, and GS-specific methods.

Method	RRE / RTE / RSE / Success	Time
HLoc (SuperPoint + SuperGlue)	2.725 / 0.099 / 0.098 / 0.756	212.3 s
FGR	157.126 / 3.328 / 0.268 / 1.000	3.4 s
REGTR	80.095 / 2.768 / 0.408 / 1.000	3.5 s
GaussReg	2.827 / 0.042 / 0.032 / 1.000	4.8 s

GaussReg is reported as 44 times faster than HLoc while maintaining comparable rotation accuracy and achieving substantially better translation and scale errors. The point-cloud baselines FGR and REGTR are markedly worse in RRE, RTE, and RSE despite their reported Success Ratio of 1.000; because the paper does not specify the success threshold, the exact relationship between these error magnitudes and the success statistic is left underdetermined.

Ablation studies clarify the role of the fine stage. The coarse-only variant (“Ours w./o. fine”) reports RRE 3.403, RTE 0.061, RSE 0.034, Success 1.000, and Time 3.7 s. Replacing the proposed fine stage with HLoc refinement yields RRE 1.104, RTE 0.186, RSE 0.278, Success 0.512, and Time 206.8 s. The fine-stage image-selection study reports that $(x,y,z,\alpha,r,g,b)$ 8 provides a good balance and that performance degrades when $(x,y,z,\alpha,r,g,b)$ 9. This pattern suggests that the benchmark is particularly sensitive to how overlapping viewpoints are selected and how photometric evidence is coupled back into 3D refinement.

6. Subsequent use, scope, and limitations

ScanNet-GSReg has subsequently served as a benchmark for automated 3D-GS sub-map alignment and fusion. “Automated 3D-GS Registration and Fusion via Skeleton Alignment and Gaussian-Adaptive Features” treats it as the primary evaluation set for a pipeline combining skeleton extraction, GA-KPConv features, GeoTransformer-based pose estimation, and multi-factor soft-threshold fusion (Liu et al., 28 Jul 2025). On the reported ScanNet-GSReg test set average, that work improves registration from RRE 3.418% to 2.595%, RTE 0.060 m to 0.045 m, and RSE 0.029% to 0.013%, while runtime increases from 4.151 s to 5.041 s. For fusion quality, the average PSNR increases from 20.270 dB to 25.212 dB, SSIM from 0.778 to 0.868, and LPIPS decreases from 0.360 to 0.305. The paper attributes these gains to anisotropy-aware feature extraction and to fusion decisions that avoid rigid thresholding in overlap regions.

The benchmark also sits within a broader ScanNet-centered GS research trajectory. LoopSplat performs online loop closure by registering 3D Gaussian splats directly on ScanNet and ScanNet++, using differentiable rendering rather than explicit correspondence search (Zhu et al., 2024). GS4 addresses ScanNet semantic SLAM with a generalizable RGB-D-to-Gaussians model and a single-iteration post-bundle-adjustment optimization step (Jiang et al., 6 Jun 2025). GTLR-GS studies LiDAR-regularized GS on ScanNet++ and explicitly frames metric-depth regularization as relevant to ScanNet-GSReg-like scenarios, although its experiments are on ScanNet++ and a custom real-world dataset rather than on the ScanNet-GSReg benchmark itself (Fang et al., 24 Mar 2026). AtlasGS, likewise in ScanNet-like indoor settings, introduces Atlanta-world-guided structural regularization for low-texture regions but is evaluated on ScanNet, ScanNet++, and Replica rather than on ScanNet-GSReg specifically (Zhang et al., 29 Oct 2025). Taken together, these works indicate that ScanNet-GSReg helped crystallize registration as a first-class problem for GS scene representations, while adjacent literature expanded toward loop closure, semantic SLAM, depth regularization, and structural priors.

The benchmark also has clear limitations. The original paper highlights a fusion limitation under differing lighting conditions between sequences, where appearance inconsistencies can emerge at fusion boundaries in the merged GS model (Chang et al., 2024). Producing each pair requires training two independent GS models for 10,000 iterations apiece, but the paper does not quantify the corresponding training time or compute cost. It also does not specify dataset file structure, GS renderer equations, or the thresholds defining successful registration. The project page is provided, but explicit dataset download links, repository URLs beyond the project page, and licensing or usage terms are not specified in the paper. These omissions matter because ScanNet-GSReg is not only a dataset but also a protocol benchmark, and protocol under-specification can complicate strict cross-paper comparability.