Cross-Instance Gaussian Splatting Registration via Geometry-Aware Feature-Guided Alignment

cs.CV Roy Amoyal, Oren Freifeld, Chaim Baskin · Mar 23, 2026
Local to this browser
What it does
The paper addresses the novel challenge of aligning independent 3D Gaussian Splatting models across different object instances within the same category—a task beyond existing same-object registration methods. The core innovation is a...
Why it matters
The core innovation is a two-stage pipeline: first, a coarse alignment using a feature-guided iterative absolute orientation solver that handles extreme initializations (180° rotations, 10× scale differences); second, a fine alignment that...
Main concern
GSA presents a technically rigorous and well-motivated solution to a genuinely novel problem in 3DGS registration. The method achieves near-perfect same-object alignment and demonstrates the first viable approach to cross-instance...
Community signal
0
0 up · 0 down
Sign in to vote with arrows
AI Review AI reviewed
Plain-language introduction

The paper addresses the novel challenge of aligning independent 3D Gaussian Splatting models across different object instances within the same category—a task beyond existing same-object registration methods. The core innovation is a two-stage pipeline: first, a coarse alignment using a feature-guided iterative absolute orientation solver that handles extreme initializations (180° rotations, 10× scale differences); second, a fine alignment that enforces multi-view feature consistency via an inverse-radiance-field formulation generalized to the similarity group $\text{Sim}(3)$. This enables the first viable category-level 3DGS registration, unlocking applications like geometrically-consistent object replacement.

Critical review
Verdict
Bottom line

GSA presents a technically rigorous and well-motivated solution to a genuinely novel problem in 3DGS registration. The method achieves near-perfect same-object alignment and demonstrates the first viable approach to cross-instance category-level registration, outperforming prior work by orders of magnitude. However, the approach's dependence on pre-trained geometry-aware features limits its applicability to categories where such features are available, and the lack of explicit code release information hinders immediate reproducibility.

“GSA's main limitation is that its performance depends on the quality of the geometry-aware features. If these are suboptimal, alignment accuracy may degrade.”
Amoyal et al. (this paper) · Section 5
“using features from [25] or [39] (instead of [21]) usually completely fails, making quantitative comparison pointless”
Amoyal et al. (this paper) · Section 4.5
What holds up

The coarse alignment stage is particularly compelling, effectively extending ICP with semantic-geometric feature constraints to solve the absolute orientation problem with unknown scale. The derivation of the multi-view feature consistency loss (Eq. 8) from inverse radiance field principles provides a solid theoretical foundation, and the ablation study validates that dropping feature guidance causes RRE to rise from near-zero to over $136^{\circ}$ on same-object tasks.

“achieving robustness to extremely poor initializations (including 180° rotation); 2) handling the case of an unknown scale; 3) enabling registration in the cross-instance setting”
Amoyal et al. (this paper) · Section 3.3
“dropping feature-based guidance in coarse alignment... Re-evaluation on Objaverse showed a drastic performance drop: coarse alignment RRE rose to 136.29°, and fine alignment to 139.82°”
Amoyal et al. (this paper) · Section 4.5
Main concerns

The primary limitation is the paper's reliance on high-quality geometry-aware features from Mariotti et al.—the authors acknowledge that substituting DINOv2 or TellingLeftfromRight features causes the method to 'usually completely fail,' making GSA inapplicable to textureless or feature-poor objects. The real-world evaluation is also limited to cars and specific CO3Dv2 categories, leaving generalization to more diverse categories unverified. Additionally, while the comparison to GaussReg is fair regarding the scale assumption, the paper does not explicitly address computational overhead relative to simpler ICP-based methods.

“using features from [25] or [39] (instead of [21]) usually completely fails”
Amoyal et al. (this paper) · Section 4.5
“The datasets capture real-world conditions, including partial observations and low-resolution, blurry images, making the task even more difficult”
Amoyal et al. (this paper) · Section 4.4
Evidence and comparison

The evidence supports the main claims: Table 2 shows an order-of-magnitude improvement over GaussReg in same-object registration, while Table 3 demonstrates the first successful results on category-level alignment with low RRE across ShapeNet categories. The comparison is fair—acknowledging that GaussReg assumes unit scale $s=1$ while GSA estimates it—and the ablation validates the contribution of each component, including the critical role of feature-based pruning in coarse alignment.

“GaussReg and DReg-NeRF assume a fixed scale ($s=1$), which is generally invalid in Novel View Synthesis... To mitigate this, its evaluation assumes a unit scale ($s=1$)”
Amoyal et al. (this paper) · Section 4.2
“GSA substantially improves alignment, achieving an order-of-magnitude gain over competing methods”
Amoyal et al. (this paper) · Section 4.2
Reproducibility

The paper provides detailed hyperparameters ($\tau_f=0.01$, learning rate $0.01$, 6 coarse iterations, 60 fine iterations, 3 diverse views) and hardware specifications (RTX 3090), which aids reproduction. However, there is no mention of code release or public repository in the main text, only a project webpage reference. The method depends on specific pre-trained features (Mariotti et al.) and SAM for masking, requiring these external dependencies to be available.

“Coarse registration uses $\tau_f=0.01$ for up to 6 iterations, and fine registration uses 3 diverse views, 60 iterations, and a learning rate of 0.01. All experiments were run on an NVIDIA RTX 3090 GPU”
Amoyal et al. (this paper) · Section 4.1
“We also extract object masks using SAM [17]”
Amoyal et al. (this paper) · Section 3.1
Abstract

We present Gaussian Splatting Alignment (GSA), a novel method for aligning two independent 3D Gaussian Splatting (3DGS) models via a similarity transformation (rotation, translation, and scale), even when they are of different objects in the same category (e.g., different cars). In contrast, existing methods can only align 3DGS models of the same object (e.g., the same car) and often must be given true scale as input, while we estimate it successfully. GSA leverages viewpoint-guided spherical map features to obtain robust correspondences and introduces a two-step optimization framework that aligns 3DGS models while keeping them fixed. First, we apply an iterative feature-guided absolute orientation solver as our coarse registration, which is robust to poor initialization (e.g., 180 degrees misalignment or a 10x scale gap). Next, we use a fine registration step that enforces multi-view feature consistency, inspired by inverse radiance-field formulations. The first step already achieves state-of-the-art performance, and the second further improves results. In the same-object case, GSA outperforms prior works, often by a large margin, even when the other methods are given the true scale. In the harder case of different objects in the same category, GSA vastly surpasses them, providing the first effective solution for category-level 3DGS registration and unlocking new applications. Project webpage: https://bgu-cs-vil.github.io/GSA-project/

Challenge the Review

Pick a starting point or write your own. Challenges run in the background, so you can keep reading while the AI investigates.

No challenges yet. Disagree with the review? Ask the AI to revisit a specific claim.