StructureGS: Structure-aware Gaussian Splatting for Articulated Object Reconstruction
StructureGS integrates structure-aware guidance with Gaussian Splatting for high-quality articulated object reconstruction, using OBB constraints to improve part boundaries and motion accuracy.
Key Findings
Methodology
This approach combines 3D Gaussian Splatting with explicit structural guidance via oriented bounding boxes (OBB). It introduces part fitting and contact losses based on OBBs, enforcing spatial coherence and physical plausibility. The method jointly optimizes per-part geometry, appearance, and motion parameters from multi-view RGB images, utilizing a multi-state transformation framework. The structure-aware losses leverage SAT for efficient contact evaluation, enabling the model to produce clear part boundaries and accurate joint motions. The framework is robust to sparse views and complex articulations, outperforming prior methods like ScrewSplat and ArticulatedGS in geometric accuracy and motion estimation.
Key Results
- On the PARIS and DTA datasets, the structure-guided model reduces Chamfer-L1 distance (CD) by over 50%, achieving 1.558 on full-view tests and 1.957 on sparse views, with joint angle errors dropping below 1° in many cases, significantly outperforming baseline methods.
- The model maintains high-quality reconstructions with sharp boundaries and physically plausible contacts, even under limited observations, demonstrating robustness and generalization.
- Ablation studies confirm that the OBB-based structure losses are critical for boundary clarity and connection accuracy, with noticeable improvements over photometry-only approaches.
Significance
This work addresses fundamental limitations in existing articulated object reconstruction methods that rely solely on photometric supervision. By explicitly modeling structural properties, it disentangles geometry and motion, leading to more accurate, interpretable, and physically consistent models. This advancement benefits applications in robotics, AR/VR, and industrial design, where precise part segmentation and motion understanding are essential. The integration of structure-aware constraints into Gaussian Splatting opens new avenues for scalable, high-fidelity 3D modeling of complex articulated objects, bridging the gap between visual realism and physical plausibility.
Technical Contribution
The key innovation is the integration of OBB-based structure-aware losses into a Gaussian Splatting framework, enabling explicit geometric and contact constraints. The part fitting loss enforces tight bounding box coverage, while the contact loss, computed via SAT, ensures physically plausible connections. The multi-state optimization jointly refines geometry, appearance, and motion, supported by extent regularization for adaptive scaling. This approach differs from prior methods by explicitly modeling structural properties rather than relying solely on photometric cues, leading to superior boundary clarity and motion accuracy.
Novelty
This is the first work to embed explicit OBB-based structural constraints into a Gaussian Splatting framework for articulated objects. Unlike previous approaches that depend on deformation or spectral clustering, this method directly encodes part boundaries and contact relationships, significantly improving geometric clarity and physical plausibility. The combination of structure-aware losses with multi-state Gaussian optimization represents a novel paradigm in 3D articulated reconstruction.
Limitations
- The method's reliance on initial OBB estimates can limit performance if the initial bounding boxes are inaccurate. Handling highly non-box-like geometries remains challenging.
- Computational complexity is high, especially with many parts and views, requiring further optimization for real-time applications.
- Current approach assumes rigid parts; extending to deformable or soft objects would require additional modeling innovations.
Future Work
Future directions include developing learning-based initialization for OBBs to improve robustness, extending the framework to handle deformable parts, and optimizing the pipeline for real-time applications. Incorporating dynamic scene modeling and exploring unsupervised structural priors could further enhance scalability and accuracy.
AI Executive Summary
Reconstructing articulated objects with multiple movable parts remains a significant challenge in 3D computer vision. Existing methods often rely solely on photometric cues, which struggle to disentangle geometry, appearance, and motion, leading to blurred boundaries and unrealistic connections. To overcome these limitations, this paper introduces StructureGS, a novel framework that integrates explicit structural guidance into 3D Gaussian Splatting. The core idea is to leverage oriented bounding boxes (OBB) to encode each part's spatial extent and contact relationships, enforcing geometric compactness and physical plausibility through dedicated loss functions.
The approach involves representing each part as a set of 3D Gaussians in a canonical space, with transformations to initial and articulated states. By optimizing these parameters jointly with OBB constraints, the method achieves sharp part boundaries and accurate motion estimation. The structure-aware losses, including part fitting and contact losses based on the Separating Axis Theorem (SAT), ensure that parts remain spatially coherent and physically connected.
Experimental results on datasets like PARIS and DTA demonstrate that StructureGS surpasses prior methods such as ScrewSplat and ArticulatedGS, reducing geometric errors by over 50% and achieving near-perfect joint motion estimates. The model maintains robustness under sparse views and complex articulations, making it suitable for real-world applications like robotic manipulation, virtual reality, and industrial design.
Overall, this work marks a significant step forward in articulated object reconstruction, combining high-fidelity rendering with explicit structural modeling. Its ability to produce clear boundaries, plausible contacts, and accurate motions opens new possibilities for automated 3D modeling of complex objects. Future work aims to extend the framework to deformable parts, real-time scenarios, and larger-scale scenes, further broadening its impact across industry and research.
Deep Dive
Abstract
Reconstructing articulated objects with multiple movable parts is essential for understanding object structure and enabling physical interaction. However, this reconstruction task poses significant challenges due to the entanglement of geometry, appearance, and motion parameters during optimization. Existing methods rely primarily on photometric supervision, which commonly fails to disentangle these interdependent components, resulting in poor part decomposition with blurred boundaries and geometric artifacts. To address this limitation, we introduce StructureGS, a reconstruction framework for articulated objects that integrates structure-aware guidance into 3D Gaussian Splatting. Our approach leverages oriented bounding boxes of object parts to enforce two key structural properties: spatial coherence, which constrains each part's geometry to remain compact and spatially coherent within its designated region, and structural connectivity, which enforces physically plausible contact relationships between adjacent parts. These properties are realized through structure-aware losses that inject explicit structural constraints into the optimization process. Extensive experiments demonstrate that our method achieves state-of-the-art performance in articulated object reconstruction, producing high-quality results with well-defined part geometries.