Orientation-anchored Gaussian Splatting (OriGS)

• OriGS is a 4D reconstruction framework that recovers dynamic 3D scenes from a single, casually captured monocular video, addressing challenges like occlusion and non-rigid deformation.
• It introduces a Global Orientation Field to propagate principal motion directions and an Orientation-aware Hyper-Gaussian with conditioned slicing for local, context-sensitive deformation.
• Empirical results show state-of-the-art performance with metrics such as PSNR 19.69, SSIM 0.716, and LPIPS 0.256, particularly excelling in complex scenes with fast rotations.

Orientation-anchored Gaussian Splatting (OriGS) is a framework for high-quality 4D reconstruction from casually captured monocular videos. It targets the recovery of a time-varying 3D scene—geometry, appearance, and motion—from a single handheld video under unconstrained, real-world dynamics, and produces a dynamic Gaussian representation that can be rendered from novel viewpoints at arbitrary times (Wu et al., 27 Sep 2025). OriGS is built around two coupled components: a Global Orientation Field that propagates principal forward directions across space and time, and an Orientation-aware Hyper-Gaussian that embeds time, space, geometry, and orientation into a unified probabilistic state, from which local deformation is inferred by Gaussian conditioning, termed conditioned slicing (Wu et al., 27 Sep 2025). Although OriGS is not named in the Gaussian Splatting review, the review’s treatment of covariance factorization, anisotropy, quaternion-based orientation, and structured dynamic Gaussian models provides the immediate mathematical and conceptual background for the method (Dalal et al., 2024).

1. Problem formulation and scope

OriGS addresses 4D reconstruction from a single, casually captured video, with the explicit aim of handling monocular input, handheld camera motion, occlusions, fast motion, non-rigid deformation, and region-specific dynamics (Wu et al., 27 Sep 2025). In this setting, the reconstruction problem is severely under-constrained: depth is ambiguous from a single camera, tracking can fail under occlusion and fast motion, and different parts of an object or scene can move differently over time (Wu et al., 27 Sep 2025).

The method is positioned against several dynamic Gaussian Splatting families already used for dynamic scene modeling: low-rank motion bases, deformation graphs or motion scaffolds, spline trajectories, and per-Gaussian MLP deformations (Wu et al., 27 Sep 2025). The stated limitation of these approaches is that they implicitly assume that motion is low-rank or smooth, and that nearby regions have similar dynamics. In simple scenes this is adequate, but in casual videos with articulated, high-frequency, and region-specific motion it leads to spatial drift, structural fragmentation, and temporal inconsistency (Wu et al., 27 Sep 2025).

OriGS reformulates the dynamic modeling problem around orientation. Its central premise is that orientation—how regions are oriented and “pointing”—can act as a global dynamic anchor across space and time, while local geometric change can be recovered conditionally from a higher-dimensional Gaussian state space (Wu et al., 27 Sep 2025). This suggests a shift away from globally shared motion bases toward a representation in which each primitive has its own learned deformation statistics, conditioned on time and local orientation.

2. Position within the Gaussian Splatting formalism

The immediate background of OriGS is the standard 3D Gaussian Splatting representation surveyed in the review literature (Dalal et al., 2024). In static 3D-GS, a scene is represented as a collection of Gaussian primitives

$G=\{\mu,\Sigma,c,o\},$

where $\mu \in \mathbb{R}^3$ is the mean, $\Sigma \in \mathbb{R}^{3\times 3}$ the covariance, $c \in \mathbb{R}^3$ the color, and $o \in \mathbb{R}$ the opacity (Dalal et al., 2024). The camera is parameterized by extrinsics $T_{\textrm{cw}}$ and intrinsics $(f_x,f_y,c_x,c_y)$, and each 3D Gaussian is projected to a 2D Gaussian in pixel space (Dalal et al., 2024).

The review identifies covariance factorization as the key orientation-related mechanism: $\Sigma = R S S^\top R^\top,$ with $R \in SO(3)$ derived from a quaternion $q$, and $\mu \in \mathbb{R}^3$0 a diagonal scale matrix (Dalal et al., 2024). In this parameterization, $\mu \in \mathbb{R}^3$1 controls the orientation of the principal axes and $\mu \in \mathbb{R}^3$2 controls anisotropic radii. The projected covariance is obtained through the Jacobian $\mu \in \mathbb{R}^3$3 of the perspective transform: $\mu \in \mathbb{R}^3$4 so orientation affects rendering both through the Gaussian’s local frame and through camera-space projection (Dalal et al., 2024). Pixel colors are then accumulated by alpha compositing in front-to-back order: $\mu \in \mathbb{R}^3$5 with $\mu \in \mathbb{R}^3$6 determined by the 2D Gaussian footprint and learned opacity (Dalal et al., 2024).

This formulation matters for OriGS because the review already isolates the same structural elements on which OriGS builds: quaternion-parameterized orientation, anisotropy, covariance conditioning, and the role of orientation in both geometric fidelity and rendering stability (Dalal et al., 2024). The review also places dynamic and structured variants—such as deformable 3DGS, sparse-controlled Gaussian Splatting, scaffolded Gaussians, and surface-aligned methods—along a continuum in which Gaussian positions and local frames are increasingly tied to anchors, templates, or motion priors (Dalal et al., 2024). OriGS can therefore be situated as a dynamic, explicitly orientation-anchored extension of this broader line of work.

3. Global Orientation Field

The first core component of OriGS is the Global Orientation Field (GOF), a field of orientation anchors defined throughout the scene and over time, intended to capture principal forward directions of motion (Wu et al., 27 Sep 2025). The GOF is estimated from long-range 3D point trajectories obtained by combining tracking, monocular depth, and camera pose estimation.

Given a monocular video with $\mu \in \mathbb{R}^3$7 frames, OriGS first constructs 2D long-range tracks $\mu \in \mathbb{R}^3$8, depth maps $\mu \in \mathbb{R}^3$9, and camera poses $\Sigma \in \mathbb{R}^{3\times 3}$0 (Wu et al., 27 Sep 2025). Each track is lifted to 3D via

$\Sigma \in \mathbb{R}^{3\times 3}$1

yielding a 3D trajectory $\Sigma \in \mathbb{R}^{3\times 3}$2 for each tracked point (Wu et al., 27 Sep 2025). These trajectories define the initial anchor positions.

Initial orientations are extracted from the first $\Sigma \in \mathbb{R}^{3\times 3}$3 frames by PCA on the centered trajectory segment. With

$\Sigma \in \mathbb{R}^{3\times 3}$4

and covariance

$\Sigma \in \mathbb{R}^{3\times 3}$5

PCA yields eigenvectors $\Sigma \in \mathbb{R}^{3\times 3}$6, and the leading eigenvector $\Sigma \in \mathbb{R}^{3\times 3}$7 is taken as the principal forward direction (Wu et al., 27 Sep 2025). This direction is embedded into an initial orientation $\Sigma \in \mathbb{R}^{3\times 3}$8.

OriGS then propagates this initial orientation across time by localized orthogonal Procrustes alignment. For a neighborhood of $\Sigma \in \mathbb{R}^{3\times 3}$9 trajectories at time $c \in \mathbb{R}^3$0 and time $c \in \mathbb{R}^3$1, it solves

$c \in \mathbb{R}^3$2

whose optimum is obtained by SVD, with

$c \in \mathbb{R}^3$3

and determinant correction to enforce $c \in \mathbb{R}^3$4 (Wu et al., 27 Sep 2025). The propagated orientation is then

$c \in \mathbb{R}^3$5

The resulting pairs $c \in \mathbb{R}^3$6 over all anchors define the GOF (Wu et al., 27 Sep 2025). Relative anchor motion between times $c \in \mathbb{R}^3$7 and $c \in \mathbb{R}^3$8 is expressed in $c \in \mathbb{R}^3$9 as

$o \in \mathbb{R}$0

which is later converted into dual quaternions for deformation blending (Wu et al., 27 Sep 2025). The significance of the GOF is that it supplies a scene-wide structural signal derived from observed motion rather than from purely abstract latent bases.

4. Orientation-aware Hyper-Gaussian and conditioned slicing

The second core component is the Orientation-aware Hyper-Gaussian, which extends each primitive from a 3D Gaussian in Euclidean space to a Gaussian on a higher-dimensional dynamic state: $o \in \mathbb{R}$1 where $o \in \mathbb{R}$2 is a position offset, $o \in \mathbb{R}$3 is a geometry offset, $o \in \mathbb{R}$4 is time, and $o \in \mathbb{R}$5 is local orientation (Wu et al., 27 Sep 2025). The geometry offset comprises scale and rotation deformation, so $o \in \mathbb{R}$6 can be regarded as $o \in \mathbb{R}$7 in the formulation described in the paper (Wu et al., 27 Sep 2025).

Each state is modeled as Gaussian: $o \in \mathbb{R}$8 with canonical mean

$o \in \mathbb{R}$9

and block covariance

$T_{\textrm{cw}}$0

The covariance is parameterized through a Cholesky factorization,

$T_{\textrm{cw}}$1

and, for efficiency, OriGS does not explicitly store the full dense matrix; instead it factors the covariance into a Cholesky-parameterized marginal on $T_{\textrm{cw}}$2 and a learnable cross-covariance with $T_{\textrm{cw}}$3 (Wu et al., 27 Sep 2025).

At a query time $T_{\textrm{cw}}$4 and local orientation $T_{\textrm{cw}}$5, OriGS computes the conditional distribution

$T_{\textrm{cw}}$6

with conditional mean

$T_{\textrm{cw}}$7

and conditional covariance

$T_{\textrm{cw}}$8

Here $T_{\textrm{cw}}$9 denotes the relative rotation operator in $(f_x,f_y,c_x,c_y)$0 (Wu et al., 27 Sep 2025).

OriGS does not sample from this conditional. Instead, it uses the conditional mean,

$(f_x,f_y,c_x,c_y)$1

which the paper calls conditioned slicing (Wu et al., 27 Sep 2025). The significance of this construction is explicit in the formulation: cross-covariance terms learn how local geometry changes when time and orientation change, enabling region-specific and context-sensitive deformation without imposing a single global low-rank basis.

5. Deformation, rendering, and optimization

OriGS starts from classical 3DGS primitives with center $(f_x,f_y,c_x,c_y)$2, rotation $(f_x,f_y,c_x,c_y)$3, scale $(f_x,f_y,c_x,c_y)$4, opacity $(f_x,f_y,c_x,c_y)$5, and color $(f_x,f_y,c_x,c_y)$6, and attaches a hyper-Gaussian to each primitive (Wu et al., 27 Sep 2025). Dynamic deformation is then produced by combining anchor-guided motion transfer with conditioned slicing.

Each Gaussian is attached to nearby oriented anchors with skinning weights $(f_x,f_y,c_x,c_y)$7 based on spatial proximity (Wu et al., 27 Sep 2025). For anchor $(f_x,f_y,c_x,c_y)$8, the relative transform from time $(f_x,f_y,c_x,c_y)$9 to $\Sigma = R S S^\top R^\top,$0 is

$\Sigma = R S S^\top R^\top,$1

which is converted to a dual quaternion $\Sigma = R S S^\top R^\top,$2 and blended over the $\Sigma = R S S^\top R^\top,$3 nearest anchors: $\Sigma = R S S^\top R^\top,$4 Applying this blended transform yields

$\Sigma = R S S^\top R^\top,$5

The local orientation at the Gaussian location is obtained by interpolation in $\Sigma = R S S^\top R^\top,$6: $\Sigma = R S S^\top R^\top,$7 (Wu et al., 27 Sep 2025).

The conditional mean is then decomposed into position, scale, and rotation offsets,

$\Sigma = R S S^\top R^\top,$8

which modulate the Gaussian: $\Sigma = R S S^\top R^\top,$9 Opacity is further modulated by temporal and orientational distance from the canonical state: $R \in SO(3)$0 so Gaussians fade out away from their learned support in $R \in SO(3)$1 space (Wu et al., 27 Sep 2025).

Rendering remains within the 3DGS rasterization pipeline: Gaussians are projected into the image plane and alpha-blended front-to-back (Dalal et al., 2024, Wu et al., 27 Sep 2025). Training is end-to-end with gradient descent and includes a photometric loss,

$R \in SO(3)$2

or a similar combination of $R \in SO(3)$3 and SSIM, a 2D correspondence loss tied to long-range tracks and depth priors, and As-rigid-as-possible (ARAP) regularization on anchor deformations (Wu et al., 27 Sep 2025). The implementation prunes Gaussians if opacity remains low and densifies regions with large photometric residuals. Reported training time is roughly 0.5–2 hours on a single RTX A6000, depending on scene length and complexity (Wu et al., 27 Sep 2025).

6. Empirical behavior, relation to prior methods, and limitations

OriGS is compared against low-rank or basis-motion methods, deformation-graph or motion-scaffold methods, and time-embedded Gaussian models such as 4DGS and 4D-rotor GS (Wu et al., 27 Sep 2025). The stated distinction is threefold: it uses a GOF extracted from scene motion rather than abstract bases, it represents dynamics through a hyper-Gaussian in $R \in SO(3)$4, and it infers deformation through Gaussian conditioning rather than directly prescribing a transform $R \in SO(3)$5 (Wu et al., 27 Sep 2025). In that sense, the per-primitive covariance acts as an implicit motion model.

Quantitative evaluation on DyCheck uses seven scenes—Apple, Block, Paper Windmill, Space Out, Spin, Teddy, and Wheel—and reports PSNR, SSIM, and LPIPS on reference viewpoints (Wu et al., 27 Sep 2025). With known poses (spc), OriGS achieves state-of-the-art average performance with Average PSNR 19.69, Average SSIM 0.716, and Average LPIPS 0.256 (Wu et al., 27 Sep 2025). It is reported to excel particularly on scenes with strong rotations and non-rigid dynamics; for Paper Windmill, the reported scores are PSNR 22.46, SSIM 0.751, and LPIPS 0.152, best across all metrics (Wu et al., 27 Sep 2025).

The ablation on DyCheck Apple is especially explicit. The paper reports: 3DGS-MLP with PSNR 13.47, SSIM 0.532, LPIPS 0.586; Deform w/ GOF with PSNR 16.28, SSIM 0.694, LPIPS 0.476; Hyper-Gaussian w/ $R \in SO(3)$6 only with PSNR 18.71, SSIM 0.750, LPIPS 0.393; and Full OriGS with PSNR 19.46, SSIM 0.807, LPIPS 0.341 (Wu et al., 27 Sep 2025). On point tracking for DyCheck, OriGS reports PCK-T 0.851, compared with CoTracker 0.803, Marbles 0.806, and MoSca 0.824 (Wu et al., 27 Sep 2025). Qualitatively, the paper states that OriGS better preserves thin spinning blades on Paper Windmill, reduces ghosting and preserves hand-object contact boundaries on Block, and maintains coherence for fast, heavily occluded motion on DAVIS Libby (Wu et al., 27 Sep 2025).

The reported limitations are equally specific. In scenes with very simple, near-rigid motion, orientation changes little, so orientation conditioning offers limited gain while still increasing computational complexity (Wu et al., 27 Sep 2025). Reconstruction quality depends on 2D tracking and depth priors; severe tracking or depth failures can degrade the orientation field and reconstruction (Wu et al., 27 Sep 2025). Orientation embedding and covariance parameterization also introduce a memory and time overhead relative to plain 3DGS (Wu et al., 27 Sep 2025).

Within the broader Gaussian Splatting landscape, the review identifies several precursors to this design: covariance factorization with explicit quaternion orientation; dynamic formulations with time-dependent rotations; ARAP-style regularization; surface-aligned Gaussians; sparse control-point motion fields; and anchored or scaffolded Gaussians (Dalal et al., 2024). This suggests that OriGS is best understood not as a departure from Gaussian Splatting, but as an overview of orientation-aware covariance modeling, anchored dynamic deformation, and probabilistic conditioning into a single 4D reconstruction framework.