Feed-forward Gaussian reconstruction from sparse posed monocular frames is hard because the input observations differ in both viewpoint and time. A new arXiv paper argues that existing methods tend to predict Gaussian primitives from observation-aligned features, which makes it difficult to reason jointly across those two dimensions.
The proposed method, HiQGS, uses hierarchical spatiotemporal querying to address this limitation. Instead of treating each observation in isolation, it organizes queries at multiple levels so that information from different viewpoints and time steps can be combined before the Gaussian parameters are predicted.
The paper is available as arXiv:2610.27604. Since it is a single source, there are no conflicting findings to compare here; the abstract presents the method as a direct response to shortcomings in existing feed-forward reconstruction approaches.