论文概要 (Paper Overview)
Field: Computer Vision Authors: Jia-Chen Zhao, Beiqi Chen, Xinyang Chen Published: 2026-06-26 arXiv: 2606.28321
Introduction
We present StructSplat, a feed-forward and generalizable 3D Gaussian reconstruction framework that operates directly on uncalibrated images without requiring camera parameters. Existing methods either rely on per-scene optimization or assume known camera poses, and often entangle geometry and appearance within a unified backbone, limiting reconstruction fidelity and generalization.
Key Idea
Our key idea is to adopt a structured representation that organizes geometry, semantic, and texture cues with explicit roles in the reconstruction process:
- Pixel-aligned feature injection — enables accurate texture modeling from 2D observations.
- Semantic-aware priors — improve global consistency across the scene.
- Camera alignment strategy — prevents information leakage and improves generalization.
- DL3DV: 28.045 PSNR, exceeding AnySplat (22.377) by +5.67 dB.
- ACID (cross-dataset): +1.94 dB over AnySplat.
- RealEstate10K (cross-dataset): +1.72 dB over AnySplat.
Results
Experiments show that StructSplat significantly outperforms prior methods on challenging benchmarks:
*Auto-collected on 2026-06-30.*