Split4D: Decomposed 4D Scene Reconstruction Without Video Segmentation
Paper#Computer Vision, 4D Scene Reconstruction🔬 Research|Analyzed: Jan 3, 2026 19:39•
Published: Dec 28, 2025 02:37
•1 min read
•ArXivAnalysis
This paper tackles the challenge of 4D scene reconstruction by avoiding reliance on unstable video segmentation. It introduces Freetime FeatureGS and a streaming feature learning strategy to improve reconstruction accuracy. The core innovation lies in using Gaussian primitives with learnable features and motion, coupled with a contrastive loss and temporal feature propagation, to achieve 4D segmentation and superior reconstruction results.
Key Takeaways
- •Proposes a novel approach to 4D scene reconstruction that avoids the instability of video segmentation.
- •Introduces Freetime FeatureGS, a new representation using Gaussian primitives with learnable features and motion.
- •Employs a streaming feature learning strategy to propagate features over time, improving reconstruction quality.
- •Achieves superior reconstruction results compared to existing methods.
Reference / Citation
View Original"The key idea is to represent the decomposed 4D scene with the Freetime FeatureGS and design a streaming feature learning strategy to accurately recover it from per-image segmentation maps, eliminating the need for video segmentation."