OCLDA-GS: Object-Centric Local Deformation-Aware Gaussian Streaming
Authors: Ashim Dahal, Nikesh Subedi, Rabab Abdelfattah, Nick Rahimi
arXiv: 2611.23416 [cs.CV]
Submitted: 17 November 2026
Subject: Computer Vision and Pattern Recognition (cs.CV)
Abstract
Streaming 3D Gaussian Splatting (3DGS) enables online reconstruction of dynamic scenes from multi-view videos for Free Viewpoint Video (FVV). Current time- and storage-efficient methods rely on selecting a subset of anchor Gaussians and performing neighborhood interpolation to achieve fast streamable 3DGS. However, these approaches can overlook object-specific motion and introduce inconsistencies near object boundaries.
We propose OCLDA-GS, an object-centric streaming framework that dynamically reallocates object- and motion-aware Gaussian anchors. This allocates greater anchor Gaussian support to objects undergoing large motion while restricting the Gaussian neighborhood within object groups to produce more precise object boundaries. In addition, the framework estimates object-wise rigid transformations (e.g., a moving hand) and leverages these estimates to learn per-anchor local deformation corrections for small deformations (e.g., wrinkles on a shirt).
We conduct experiments on three dynamic multi-view video datasets and demonstrate improvements over previous baselines, alongside improved quality between refinement timesteps. Under matched refinement budgets, our method improves PSNR over Instant Gaussian Stream (IGS) by 0.169 dB on N3DV, 0.202 dB on Meeting Room, and 0.199 dB on Google Immersive, while reducing the mean PSNR drop between refinement timesteps by 9.2% and 13.8% on N3DV and Meeting Room, respectively.
Key Contributions
- Object-centric anchor allocation: Dynamically reallocates object- and motion-aware Gaussian anchors, providing additional support for objects undergoing large motion.
- Boundary-aware neighborhood restriction: Constrains Gaussian neighborhoods within object groups to improve precision at object boundaries.
- Hybrid deformation modeling: Combines object-wise rigid transformation estimates with learned per-anchor local deformation corrections to capture both large-scale motion and fine surface detail.
- Improved temporal stability: Reduces quality degradation between refinement timesteps, supporting smoother streaming reconstruction.
Experimental Results
Evaluated on three dynamic multi-view video datasets under matched refinement budgets:
| Dataset | PSNR Gain over IGS | Mean PSNR Drop Between Timesteps |
|---|---|---|
| N3DV | +0.169 dB | β9.2% |
| Meeting Room | +0.202 dB | β13.8% |
| Google Immersive | +0.199 dB | β |
Metadata
- Cite as: arXiv:2611.23416 [cs.CV]
- DOI: https://doi.org/10.48550/arXiv.2611.23416 (arXiv-issued via DataCite)
- Version: arXiv:2611.23416v1
- Submission history: From Ashim Dahal, submitted 17 November 2026
via ArXiv CV
