arXiv:2512.04021v1 Announce Type: new 
Abstract: Reconstructing and understanding 3D scenes from unposed sparse views in a feed-forward manner remains as a challenging task in 3D computer vision. Recent approaches use per-pixel 3D Gaussian Splatting for reconstruction, followed by a 2D-to-3D feature lifting stage for scene understanding. However, they generate excessive redundant Gaussians, causing high memory overhead and sub-optimal multi-view feature aggregation, leading to degraded novel view synthesis and scene understanding performance. We propose C3G, a novel feed-forward framework that estimates compact 3D Gaussians only at essential spatial locations, minimizing redundancy while enabling effective feature lifting. We introduce learnable tokens that aggregate multi-view features through self-attention to guide Gaussian generation, ensuring each Gaussian integrates relevant visual features across views. We then exploit the learned attention patterns for Gaussian decoding to efficiently lift features. Extensive experiments on pose-free novel view synthesis, 3D open-vocabulary segmentation, and view-invariant feature aggregation demonstrate our approach's effectiveness. Results show that a compact yet geometrically meaningful representation is sufficient for high-quality scene reconstruction and understanding, achieving superior memory efficiency and feature fidelity compared to existing methods.

تم تقديم إطار C3G كنهج جديد لتعلم تمثيلات ثلاثية الأبعاد مضغوطة باستخدام 2K Gaussian، مما يعالج التحديات في إعادة بناء المشاهد ثلاثية الأبعاد من وجهات نظر متناثرة غير موضوعة. تقلل هذه الطريقة من التكرار في توليد Gaussian بينما تعزز رفع الميزات من خلال رموز قابلة للتعلم تجمع ميزات متعددة الرؤى عبر الانتباه الذاتي.

Se ha introducido el marco C3G como un enfoque novedoso para aprender representaciones 3D compactas utilizando 2K Gaussianos, abordando los desafíos en la reconstrucción de escenas 3D a partir de vistas dispersas no posicionadas. Este método minimiza la redundancia en la generación de Gaussianos mientras mejora el levantamiento de características a través de tokens aprendibles que agregan características de múltiples vistas mediante autoatención.

Le cadre C3G a été introduit comme une nouvelle approche pour apprendre des représentations 3D compactes en utilisant 2K Gaussiens, répondant aux défis de la reconstruction de scènes 3D à partir de vues éparses non posées. Cette méthode minimise la redondance dans la génération de Gaussiens tout en améliorant le levage des caractéristiques grâce à des jetons apprenables qui agrègent les caractéristiques multi-vues via l'auto-attention.

The C3G framework has been introduced as a novel approach to learning compact 3D representations using 2K Gaussians, addressing challenges in 3D scene reconstruction from unposed sparse views. This method minimizes redundancy in Gaussian generation while enhancing feature lifting through learnable tokens that aggregate multi-view features via self-attention.

C3G: Learning Compact 3D Representations with 2K Gaussians

Was this article worth reading? Share it

LucidQuery AI

Deptho.ai

Artefacts.ai

The Visualizer

4o Image Gen

Blunge

Ready to build your own newsroom?