Gaussian Masked Autoencoders

2025-01-06

AI Summary

Existing image representation learning methods lack sufficient spatial understanding, particularly regarding geometric structure and 3D-aware reasoning. Method: This paper proposes embedding differentiable 3D Gaussian splatting into the Masked Autoencoder (MAE) framework—yielding the first differentiable Gaussian rendering architecture for general-purpose image representation learning. It jointly reconstructs images end-to-end in pixel space while implicitly modeling and optimizing an intermediate 3D Gaussian representation, enabling differentiable rendering and co-optimization of semantic features with spatial structures (e.g., depth, hierarchy, edges). Contribution/Results: We introduce Gaussian primitives to self-supervised image representation learning for the first time. Our method achieves MAE-level high-level semantic quality while enabling zero-shot spatially aware tasks—including foreground-background segmentation, image layering, and edge detection—without task-specific supervision. This work establishes a new paradigm for high-fidelity, geometry-aware visual representation learning.

Technology Category

• Computer Vision: Representation Learning for Vision
• Machine Learning: Representation Learning
• Knowledge Representation and Reasoning: Geometric, Spatial, and Temporal Reasoning

Application Category

• Search and Retrieval-Augmented AI: Web query analysis, representation and understanding
• Graph Algorithms and Modeling for the Web: Graph embeddings and representation learning for Web-related graphs
• Semantics and Knowledge: Methods to enhance, augment, integrate or synergize semantic models such as knowledge graphs and LLMs

Abstract

This paper explores Masked Autoencoders (MAE) with Gaussian Splatting. While reconstructive self-supervised learning frameworks such as MAE learns good semantic abstractions, it is not trained for explicit spatial awareness. Our approach, named Gaussian Masked Autoencoder, or GMAE, aims to learn semantic abstractions and spatial understanding jointly. Like MAE, it reconstructs the image end-to-end in the pixel space, but beyond MAE, it also introduces an intermediate, 3D Gaussian-based representation and renders images via splatting. We show that GMAE can enable various zero-shot learning capabilities of spatial understanding (e.g., figure-ground segmentation, image layering, edge detection, etc.) while preserving the high-level semantics of self-supervised representation quality from MAE. To our knowledge, we are the first to employ Gaussian primitives in an image representation learning framework beyond optimization-based single-scene reconstructions. We believe GMAE will inspire further research in this direction and contribute to developing next-generation techniques for modeling high-fidelity visual data. More details at https://brjathu.github.io/gmae

Problem

• Image Depth Perception
• Spatial Position Learning
• Visual Recognition

Innovation

• GMAE
• 3D Techniques
• Image Depth Understanding