Temporal Gaussian Hierarchy

Temporal Gaussian Hierarchy is a research-focused framework for high-fidelity 4D scene reconstruction and long-term dynamic view synthesis from monocular videos. Built on a hierarchical organization of 3D Gaussians, the method models complex, time-varying environments by decomposing motion and appearance across multiple temporal scales. The system can ingest long, casually captured videos and recover consistent geometry, appearance, and motion, enabling free-viewpoint rendering over extended time spans. Its hierarchical temporal structure allows it to handle large motion, recurring dynamics, and long-term changes while maintaining spatial detail and temporal coherence.

Key capabilities include reconstructing dynamic scenes with moving objects and deforming surfaces, generating novel views at arbitrary time steps, and maintaining global consistency across long sequences. The framework supports detailed per-frame rendering as well as temporally aggregated representations that capture stable scene components. Typical use cases include dynamic scene capture for virtual and augmented reality, performance and motion analysis, content creation for film and animation, and research on 4D scene understanding. By leveraging a temporally structured Gaussian representation, Temporal Gaussian Hierarchy addresses challenges of drift, occlusion, and motion blur that affect traditional video-based reconstruction methods, providing a robust foundation for long-term, high-quality dynamic view synthesis.