Spiky 3D Gaussian Cutter (SGC)

• The Spiky 3D Gaussian Cutter (SGC) is a post-processing module that refines interactive 3D segmentation by geometrically adjusting Gaussians to align with SAM mask boundaries.
• It operates after the preliminary segmentation, using projected covariance, eigen-decomposition, and coverage ratios to shift and scale Gaussians for smoother boundaries.
• Empirical results demonstrate that SGC improves segmentation quality by increasing IoU and accuracy while reducing spurious protrusions from elongated Gaussian supports.

Seg-Wild introduces the Spiky 3D Gaussian Cutter (SGC) as a post-processing module for interactive segmentation in 3D Gaussian Splatting under unconstrained image collections. Within the Seg-Wild pipeline, SGC operates after a preliminary 3D segmentation mask has been produced from per-Gaussian affinity features and a user-prompted SAM mask, and its purpose is to smooth segmentation boundaries by removing protruding parts of anisotropic 3D Gaussians while retaining useful interior contributions (Bao et al., 10 Jul 2025). In the formulation reported for Seg-Wild, SGC does not retrain features or alter the selection criterion that defines the segmented Gaussian set; instead, it adjusts Gaussian geometry by estimating how much of each projected long axis remains inside the 2D SAM mask and then updating center and scale accordingly (Bao et al., 10 Jul 2025).

1. Position within the Seg-Wild pipeline

In Seg-Wild, interactive segmentation proceeds in four stages: reconstruction and feature-field creation with a GS-W backbone, SAM-guided feature compactness using the Scale-Adaptive Segmentation Module (SASM), user-driven interactive selection, and finally SGC as the last stage of the segmentation process (Bao et al., 10 Jul 2025). The scene is reconstructed with 3D Gaussian Splatting, and each 3D Gaussian is augmented with an affinity feature vector trained from SAM features. SASM then uses SAM masks to make affinity features compact and object-consistent across views. A user click on an object in a reference image defines a target whose rendered feature embedding is compared to the per-Gaussian affinity features, yielding a preliminary 3D segmentation mask $M_{3D}$. SGC subsequently refines this mask geometrically (Bao et al., 10 Jul 2025).

The motivation for SGC follows directly from the properties of 3D Gaussian Splatting. The preliminary mask $M_{3D}$ is defined by feature similarity together with a 2D SAM mask $M_{2D}$. However, 3D Gaussians are generally anisotropic ellipsoids. As a result, some Gaussians whose centers lie inside the desired SAM region can still extend beyond the true object boundary along elongated axes. When such Gaussians are retained in the segmented set, their projected footprint produces visually noticeable sharp protrusions or spikes on the boundary. Seg-Wild describes SGC as a mechanism to smooth those abnormal 3D Gaussians rather than deleting them outright, because the interior part of a long Gaussian may still contribute meaningfully to the object (Bao et al., 10 Jul 2025).

This placement makes SGC tightly coupled to both the geometric representation of 3DGS and the interaction model of Seg-Wild. It relies on the same SAM mask used during the click interaction, but it operates on Gaussian centers, projected covariance, and scale parameters rather than on the affinity feature embeddings themselves (Bao et al., 10 Jul 2025).

2. Preliminary segmentation mask and SGC inputs

Before SGC is applied, Seg-Wild forms a rendered feature map

$$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$

and uses prompt points $$\mathscr{pp} = \{(u_i, v_i)\}_{i=1}^{N_{\mathscr{pp}}}$$ on a reference image to compute cosine similarities between the prompt feature and every Gaussian affinity feature: $$s_i = \frac{fe_{rend}(u_i, v_i) \cdot a_f^T}{\| fe_{rend}(u_i, v_i) \| \| a_f \|}, \quad s_i \in \mathbb{R}^{N_\mathcal{G}} \tag{25}$$ These similarities are fused by a softmax-weighted combination: $$s_{fus} = \sum_{i=1}^{N_{\mathscr{pp}}}\frac{e^{s_i}}{\sum_{j=1}^{N_{\mathscr{pp}}}e^{s_j}} s_i \tag{26}$$ Each Gaussian center $p_j$ is projected to the image plane by

$(u_j, v_j) = \pi_{cam}(p_j) \tag{27}$

and the preliminary segmented Gaussian set is then defined as

$$M_{3D} = \{ p_j \mid s_{fus} > \tau,\; M_{2D}(u_j, v_j) = 1 \}, \quad \tau = 0.5 \tag{28}$$

(Bao et al., 10 Jul 2025)

SGC takes this preliminary set of segmented Gaussians, denoted in the paper as $M_{3D}$0, together with each Gaussian’s 3D covariance $M_{3D}$1, center $M_{3D}$2, and the SAM mask $M_{3D}$3 used for the same interaction (Bao et al., 10 Jul 2025). Its output is a modified center $M_{3D}$4 and an updated scale $M_{3D}$5, stored through the covariance parameterization, so that the Gaussian is effectively cut back where its support extends outside the intended boundary (Bao et al., 10 Jul 2025).

A central conceptual feature is that SGC does not redefine membership in $M_{3D}$6. It accepts the feature-selected segmented set and then re-parameterizes Gaussians so that their rendered support becomes more consistent with the SAM boundary. This suggests a division of labor in Seg-Wild: feature similarity handles semantic target identification, while SGC handles boundary regularization (Bao et al., 10 Jul 2025).

3. Geometry of the cutting operation

SGC begins by projecting the Gaussian covariance into the image plane: $M_{3D}$7 where $M_{3D}$8 is the Jacobian of the projection affine approximation (Bao et al., 10 Jul 2025). From $M_{3D}$9, the method computes the maximum eigenvalue $M_{2D}$0 and corresponding eigenvector $M_{2D}$1, which define the long principal axis of the projected 2D ellipse. If the projected center is $M_{2D}$2, the endpoints of the long axis are

$M_{2D}$3

(Bao et al., 10 Jul 2025)

SGC then samples $M_{2D}$4 intermediate pixels $M_{2D}$5 along the segment joining $M_{2D}$6 and $M_{2D}$7, and computes a coverage ratio

$M_{2D}$8

where $M_{2D}$9 is 1 for pixels inside the SAM mask and 0 otherwise (Bao et al., 10 Jul 2025). The ratio $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$0 measures how much of the Gaussian’s projected long axis lies inside the user-specified 2D segmentation.

The actual “cut” is implemented by shifting the 2D center and shrinking the scale: $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$1 Afterward, $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$2 is back-projected to 3D to obtain the updated center, and the covariance is updated through the revised scale parameter (Bao et al., 10 Jul 2025).

The interpretation provided in Seg-Wild is direct. If $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$3, the long axis is entirely inside the mask, and no modification is applied. If $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$4, the Gaussian is effectively removed or made negligible because its scale goes to zero. For intermediate values, SGC both shifts and shrinks the Gaussian proportionally to the uncovered fraction, approximating retention of the part that lies inside the SAM-defined object region (Bao et al., 10 Jul 2025).

4. Geometric interpretation and boundary behavior

Seg-Wild uses the term “spiky” for Gaussians whose projected support creates thin, elongated protrusions at object boundaries. In 3DGS, each Gaussian is an anisotropic ellipsoid specified by a covariance $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$5. Spikiness therefore refers not to a separate primitive type but to a configuration in which a Gaussian has a long principal axis and extends beyond the target region although its center satisfies the segmentation criterion (Bao et al., 10 Jul 2025).

The source of the issue is the center-based nature of the preliminary mask. Equation (28) checks whether the center projects inside $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$6 and whether the affinity feature is sufficiently similar, but it does not constrain the full spatial footprint of the Gaussian. Elongated Gaussians near object boundaries can therefore be admitted even when a substantial portion of their support lies outside the desired object (Bao et al., 10 Jul 2025).

SGC modifies these cases in two coordinated ways. First, it shifts the center along the long-axis direction $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$7, thereby repositioning the Gaussian toward the in-mask region. Second, it scales the Gaussian by the coverage ratio $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$8, reducing the influence of Gaussians whose long-axis support is only weakly covered by the SAM mask (Bao et al., 10 Jul 2025). Seg-Wild emphasizes that this is not all-or-nothing pruning. The method is designed to preserve useful interior contributions that would be lost if an entire elongated Gaussian were simply deleted.

The qualitative effect reported in the paper is smoother, more coherent boundaries and removal of erroneous protruding components. Figure 1, described as the SGC ablation, shows that boundaries are jagged and contain thin protrusions without SGC, whereas the post-processed result is visually cleaner (Bao et al., 10 Jul 2025). This suggests that SGC functions as a boundary-consistency regularizer operating directly in Gaussian parameter space.

5. Interaction with features, SAM, and implementation choices

Although SGC is part of the segmentation stage, it does not operate on the learned affinity feature embeddings. Those embeddings are used upstream to determine which Gaussians belong to the target object, while SGC uses geometry and the 2D SAM mask to refine the selected set (Bao et al., 10 Jul 2025). The same mask $$fe_{rend} = \sum_{u=0}^{W-1}\sum_{v=0}^{H-1}\sum_{i=1}^{N_{\mathcal{G}}} af_i \,\alpha_i(u, v)\prod_{j=1}^{i-1}(1-\alpha_j(u,v)) \tag{9}$$9 that participates in defining $$\mathscr{pp} = \{(u_i, v_i)\}_{i=1}^{N_{\mathscr{pp}}}$$0 is reused to estimate the coverage ratio along the Gaussian’s projected long axis.

Seg-Wild explicitly presents SGC as a post-processing step rather than a training component. It is applied after a user interaction has generated a preliminary segmented set, and it is not part of the feature optimization loop (Bao et al., 10 Jul 2025). This design has two implications. First, SGC can be viewed as interaction-conditioned geometry editing rather than representation learning. Second, its behavior is specific to the reference view and prompt used in that interaction, since the paper describes SGC as using the camera of the reference view in which prompt points are provided (Bao et al., 10 Jul 2025).

The implementation details given in Seg-Wild are limited but specific where they are explicit. The threshold for selecting Gaussians into $$\mathscr{pp} = \{(u_i, v_i)\}_{i=1}^{N_{\mathscr{pp}}}$$1 is $$\mathscr{pp} = \{(u_i, v_i)\}_{i=1}^{N_{\mathscr{pp}}}$$2, with ablations reported for $$\mathscr{pp} = \{(u_i, v_i)\}_{i=1}^{N_{\mathscr{pp}}}$$3 and $$\mathscr{pp} = \{(u_i, v_i)\}_{i=1}^{N_{\mathscr{pp}}}$$4 (Bao et al., 10 Jul 2025). The number of samples $$\mathscr{pp} = \{(u_i, v_i)\}_{i=1}^{N_{\mathscr{pp}}}$$5 used along the long axis is referenced in Eq. (31) but no numeric value is specified (Bao et al., 10 Jul 2025). Likewise, the paper states that “scale” is a parameter stored in the covariance representation, following standard 3DGS parameterization, but does not further restate the decomposition into rotation and scaling within the SGC section (Bao et al., 10 Jul 2025).

A plausible implication is that SGC is lightweight relative to the learned feature field, because its required computations are local to each segmented Gaussian: covariance projection, eigendecomposition in 2D, line sampling against a binary mask, and parameter update. The paper itself, however, limits its concrete claims to the equations and the statement that SGC is a post-processing refinement (Bao et al., 10 Jul 2025).

6. Empirical effect, limitations, and relation to compact 3DGS research

Seg-Wild evaluates SGC directly in an ablation study. Table 2 compares “w/o SGC” against the complete model on three scenes: Brandenburg Gate (BG), Taj Mahal (TM), and Trevi Fountain (TF). The reported numbers are as follows (Bao et al., 10 Jul 2025):

Scene	w/o SGC	Complete model
BG	IoU 0.871, Acc 0.934	IoU 0.930, Acc 0.968
TM	IoU 0.797, Acc 0.903	IoU 0.846, Acc 0.920
TF	IoU 0.795, Acc 0.912	IoU 0.827, Acc 0.924

The paper summarizes these results by stating that removing SGC leads to an average decrease of 4.6% in IoU and 1.9% in Accuracy, indicating a consistent quantitative benefit from the pruning mechanism (Bao et al., 10 Jul 2025). Qualitatively, Seg-Wild also reports that competing methods such as SAGA struggle to remove floating and spiky 3D Gaussians, whereas Seg-Wild with SGC produces smoother and more coherent segmentation boundaries (Bao et al., 10 Jul 2025).

The main limitation explicitly associated with SGC is its dependence on SAM mask quality. In the paper’s failure-case discussion, if SAM omits relevant object parts, then the coverage ratio $$\mathscr{pp} = \{(u_i, v_i)\}_{i=1}^{N_{\mathscr{pp}}}$$6 for Gaussians representing those omitted regions will be reduced, and SGC may shrink or cut away legitimate object structure, producing under-segmentation (Bao et al., 10 Jul 2025). The paper does not explicitly discuss other possible failure modes such as sensitivity to the sampling density along the axis or over-cutting very thin details.

In a broader 3DGS context, SGC differs from compactness-oriented work such as SteepGS and adaptive pruning approaches. SteepGS analyzes density control as an optimization problem and derives principled rules for when splitting is beneficial, showing that two offspring along the minimum-eigenvalue direction of a splitting matrix are sufficient and optimal under its second-order analysis (Wang et al., 8 May 2025). That work is concerned with controlling growth and compactness during reconstruction rather than refining segmentation boundaries. Similarly, adaptive pruning methods that operate after densification and may combine pruning schedules with more expressive primitives, such as Difference-of-Gaussians representations, target reconstruction-aware compactness and model-size reduction rather than mask-conditioned boundary trimming (Wang et al., 27 Feb 2026). Relative to these lines of work, SGC is specialized: it leverages the geometric support of segmented Gaussians and a user-conditioned SAM mask to perform local boundary correction in interactive 3D segmentation (Bao et al., 10 Jul 2025).

This specialization is the distinctive feature of SGC. It is not a general density-control method, not a training-time sparsification strategy, and not a new primitive family. It is a segmentation-specific geometric post-process that estimates the in-mask support of each selected Gaussian and re-parameterizes it accordingly, thereby reducing spikes while preserving useful interior mass (Bao et al., 10 Jul 2025).