A scene can have a normalized average distance from the camera yet still place neighboring Gaussians at very different spacings. That mismatch makes the right Gaussian size harder to learn across scenes. The key insight here is to normalize the quantity that Gaussian coverage depends on directly: representative spacing between Gaussian centers.
Key Findings
- Radial normalization leaves scale ambiguity. The paper reports up to 4.5× variation in median Gaussian spacing across scene categories under radial-distance normalization, even though that normalization standardizes a global geometric statistic.
- GSNorm aligns the scale reference with Gaussian coverage. It uses the scene-level median of depth divided by the geometric mean of focal lengths as the normalization factor, setting median spacing to approximately one while preserving local variation.
- The evaluation tests more than familiar interpolated views. Its benchmark spans seven datasets, input counts from 1 to 64 views, and targets with controlled overlap to probe sparse-input and viewpoint-extrapolation behavior.
- The reported gains address cross-scene generalization. Experiments show state-of-the-art results across most settings, with particularly strong improvements when scene scale or viewpoint changes substantially.
Method and Trade-offs
The approach builds on geometry predicted by a frozen foundation model, then rescales depths, geometry, and camera translations before predicting per-pixel Gaussian attributes. This makes Gaussian scale supervision more consistent without removing meaningful local spacing differences. It is a useful design choice when feed-forward reconstructions must transfer across varied scenes, but it inherits the geometry model’s weaknesses: extremely sparse or difficult inputs can undermine geometry, and areas never seen in the input views may remain incomplete.
Who It Suits
A good fit for researchers evaluating generalizable, feed-forward 3D reconstruction across diverse scenes and extrapolated views. Look elsewhere if the task requires reliable completion of entirely unobserved regions, or if the available images are too sparse or challenging for the geometry prior to recover dependable structure.