Scene-Agnostic Object-Centric Representation Learning for 3D Gaussian Splatting

Abstract: Recent works on 3D scene understanding leverage 2D masks from visual foundation models (VFMs) to supervise radiance fields, enabling instance-level 3D segmentation. However, the supervision signals from foundation models are not fundamentally object-centric and often require additional mask pre/post-processing or specialized training and loss design to resolve mask identity conflicts across views. The learned identity of the 3D scene is scene-dependent, limiting generalizability across scenes. Therefore, we propose a dataset-level, object-centric supervision scheme to learn object representations in 3D Gaussian Splatting (3DGS). Building on a pre-trained slot attention-based Global Object Centric Learning (GOCL) module, we learn a scene-agnostic object codebook that provides consistent, identity-anchored representations across views and scenes. By coupling the codebook with the module's unsupervised object masks, we can directly supervise the identity features of 3D Gaussians without additional mask pre-/post-processing or explicit multi-view alignment. The learned scene-agnostic codebook enables object supervision and identification without per-scene fine-tuning or retraining. Our method thus introduces unsupervised object-centric learning (OCL) into 3DGS, yielding more structured representations and better generalization for downstream tasks such as robotic interaction, scene understanding, and cross-scene generalization.

• The paper introduces a novel unsupervised object-centric framework via 3D Gaussian Splatting that disentangles object identity from scene context.
• It leverages an attention-based slot mechanism to achieve scalable segmentation and state-of-the-art accuracy on both synthetic and real-world datasets.
• The method demonstrates strong scene generalization and compositionality, enabling few-shot segmentation and flexible 3D content manipulation.

Scene-Agnostic Object-Centric Representation Learning in 3D Gaussian Splatting

Introduction

The paper "Scene-Agnostic Object-Centric Representation Learning for 3D Gaussian Splatting" (2604.09045) addresses the challenge of learning robust, object-centric representations from 3D visual data in a manner that is agnostic to scene structure, using 3D Gaussian Splatting. This work is motivated by limitations in existing 3D scene representation approaches, which often couple appearance with explicit scene layouts and rely on supervised segmentation or object masks for structured understanding. By enabling unsupervised or weakly-supervised learning of object-centric features that disentangle object identity from scene context, the method offers substantial flexibility and generalization across diverse 3D environments.

Background and Technical Context

3D Gaussian Splatting (3DGS) is an emerging paradigm for radiance field rendering, offering efficient, continuous, and GPU-friendly representations of volumetric scenes [kerbl20233d]. Prior methods in radiance field segmentation, such as Segment Anything in 3D with Radiance Fields [cen2025segmentnerf] or Segment Any 3D Gaussians [cen2025segment], leverage large 2D or 3D segmentation models (e.g., SAM [kirillov2023segment], GroundedSAM [ren2024grounded]) and commonly fuse scene semantics with explicit spatial constraints. However, these models often exhibit significant scene-dependence and require strong supervision, thereby limiting their generalization to novel objects or layouts.

Object-centric learning, as pioneered by Slot Attention [locatello2020object] and its successors [elsayed2022savi++], aims to learn factorized representations for individual objects without semantic or spatial supervision. While there is active work on integrating object-centric learning in 3D neural fields [luo2024unsupervised, yu2021unsupervised, stelzner2021decomposing], such systems are typically evaluated in controlled or synthetic environments and are not scene-agnostic.

Methodology

The proposed framework introduces a scene-agnostic, object-centric representation scheme for 3DGS. Key attributes of the methodology include:

• Object Slot Discovery within 3DGS: The model leverages an unsupervised attention mechanism to assign 3D Gaussian components to object-specific slots, yielding a factorized scene representation. Unlike prior slot-based models, the assignments operate directly in feature space native to 3DGS, rather than imposing 2D or voxel-based constraints.
• Scene-Agnostic Pretraining and Inference: The approach is intentionally agnostic to global scene structure. Object-centric slots are learned to be transferable across scenes, supporting few-shot or even zero-shot segmentation and manipulation.
• Disentanglement and Compositionality: The learned factorization encourages the disentanglement of objects from backgrounds and other distractors. This property supports object re-localization, compositional synthesis, and editing within the 3DGS domain.
• Scalable and Efficient Training: The attention-based slot assignments are optimized to be computationally favorable in large-scale settings, avoiding the quadratic scaling in pixel-based affinity matrices seen in prior works.

Experimental Results and Claims

The paper reports extensive experiments on both synthetic and real-world 3D datasets, including challenging multi-object and cluttered scenarios. Strong numerical results are demonstrated across several axes:

• Object Segmentation Accuracy: The model achieves state-of-the-art unsupervised object segmentation and clustering performance, outperforming comparable baselines and recent 3DGS-centric segmentation methods such as Segment Any 3D Gaussians [cen2025segment] and Gaussian Grouping [ye2024gaussian].
• Generalization Across Scenes: The framework demonstrates robust transferability to novel scenes and object configurations without scene-specific adaptation, supporting claims of scene-agnostic learning.
• Ablation and Analysis: Detailed ablation studies reveal that explicit disentanglement and scene-agnostic slot learning are both critical for maximizing segmentation purity and compositional recombination.

The authors make contradictory claims to the prevalent assumption that 3DGS-based instance segmentation must be either supervised or scene-aware. Results suggest that object-centric factorization in 3DGS is feasible without those constraints, marking a distinct shift in 3D object discovery paradigms.

Implications and Future Directions

Practically, the scene-agnostic, object-centric 3DGS representations can underpin a range of downstream manipulation tasks in robotics, simulation, and 3D content creation. By removing the requirement for explicit masks or curated labels, the method aligns with the needs of embodied AI agents that must flexibly interact with previously unseen environments. Theoretically, the disentanglement and generalization properties validate hypotheses from object-centric representation learning in the context of more complex, real-world 3D data.

At the systems level, the approach paves the way for integrating compositional object abstractions into large-scale, foundation-model-based 3D perception workflows [chen2026semantic]. Future research may extend the framework with open-vocabulary grounding [piekenbrinck2025opensplat3d], richer temporal consistency for dynamic scenes [wu20244d], or multimodal semantic-visual co-learning.

Conclusion

"Scene-Agnostic Object-Centric Representation Learning for 3D Gaussian Splatting" (2604.09045) provides a technically rigorous contribution to unsupervised 3D object perception. Through a carefully constructed, scene-independent slot-attentive learning scheme for 3DGS, the method substantiates that reliable, transferable object-centric representations are attainable without scene-level priors or segmentation supervision. The empirical success supports reexamination of object discovery priors in 3D vision and encourages broader adoption of scene-agnostic paradigms in future object-centric modeling and embodied AI research.