Single image-based Gaussian splatting for 3D reconstruction of movable articulated objects

초록

Reconstructing articulated objects as high-fidelity three-dimensional (3D) digital replicas is a fundamental challenge in computer vision, with rapidly growing demand due to its broad applicability in robotics, digital twins, and virtual environments. Recent methods based on 3D Gaussian Splatting (3DGS) have enabled photorealistic and efficient modeling of articulated objects from multi-view images. However, such approaches still face practical limitations, including the requirement of a large number of view-specific images and precise camera pose estimation. To overcome these challenges, we present Single Image-based Gaussian splatting for 3D reconstruction of Movable Articulated objects (SIGMA), a novel framework that reconstructs photorealistic articulated objects from only a single RGB image per articulation state, without relying on ground truth articulation parameters, segmentation masks, or explicit camera pose input. To synthesize 3D Gaussians conditioned on a single image, we leverage a pre-trained 3D generative network and introduce Triplane-based Adaptive Cumulative Occupancy Scaling (TACOS), a learnable alignment module that normalizes independently generated Gaussian primitives using volumetric occupancy cues projected onto triplanes. Furthermore, we propose a convex hull-based refinement strategy to enhance geometric coherence and reduce generative inconsistencies. Our approach enables lightweight and scalable modeling of articulated objects while maintaining high rendering quality. Extensive experiments on both synthetic and real-world benchmarks demonstrate that SIGMA achieves state-of-the-art performance under the single-view protocol, outperforming previous methods that rely on multi-view images and corresponding viewpoints.

키워드