4DGS360: 360{\deg} Gaussian Reconstruction of Dynamic Objects from a Single Video

2026-07-23 · Source: cs.CV updates on arXiv.org · Field: Technology & Digital — Artificial Intelligence & Machine Learning, Robotics & Autonomous Systems · Depth: Expert, extended

Summary

4DGS360 is a diffusion-free framework for 360-degree dynamic object reconstruction from single monocular videos. It addresses limitations of existing methods that struggle with consistent 360-degree geometry due to 2D-native priors overfitting visible surfaces. The framework introduces AnchorTAP3D, an advanced 3D-native initialization tracker that leverages confident 2D track points as anchors to suppress drift and resolve depth ambiguity in occluded regions. This robust initialization, combined with optimization, yields coherent 360-degree 4D reconstructions. The authors also present iPhone360, a new benchmark dataset with test cameras up to 135 degrees apart from training views, enabling comprehensive 360-degree evaluation. 4DGS360 achieves superior performance on iPhone360, iPhone, and DAVIS datasets.

Key takeaway

For computer vision engineers developing 4D reconstruction systems from monocular video, 4DGS360 offers a robust approach to achieve complete 360-degree object geometry. You should integrate 3D-native, anchor-guided tracking for initialization to overcome occlusion challenges and ensure consistent reconstructions, especially for extreme novel viewpoints. This method improves fidelity and temporal coherence, making it suitable for VR/AR applications.

Key insights

4DGS360 enables consistent 360-degree dynamic object reconstruction from monocular video using 3D-native, occlusion-aware initialization.

Principles

• 3D-native initialization mitigates geometric ambiguity in occluded regions.
• Anchor-guided 3D tracking improves long-term reliability and resolves depth ambiguity.
• Rigidity regularization maintains geometric consistency across non-adjacent frames.

Method

AnchorTAP3D integrates 2D and 3D tracking, using high-confidence 2D tracks as anchors to guide 3D point trajectory estimation, followed by dynamic Gaussian optimization with rigidity and render-based regularization.

In practice

• Use AnchorTAP3D for robust 3D point tracking in dynamic scenes.
• Employ ARAP regularization to preserve local rigidity in 4D reconstructions.
• Evaluate 360-degree reconstruction using datasets with extreme novel viewpoints.

Topics

• 4D Gaussian Splatting
• Dynamic Object Reconstruction
• Monocular Video
• 3D Point Tracking
• AnchorTAP3D
• Novel View Synthesis
• iPhone360 Dataset

Best for: AI Scientist, Computer Vision Engineer, Research Scientist

Related on AIssential

• Hand-4DGS: Feed-Forward 3D Gaussian Splatting for 4D Hand Reconstruction from Egocentric Videos — Hand-4DGS is the first feed-forward 3D Gaussian Splatting framework for dynamic 4D hand reconstruction from egocentric videos.
• DGSfM: Depth-Guided Scale-Aware Global Structure-from-Motion — DGSfM enhances global SfM robustness and accuracy by integrating monocular depth maps and depth-aware filtering.
• SM4RT: Learning Structured Motion Geometry for 4D Reconstruction — SM4RT reconstructs 4D dynamic scenes by modeling motion as structured rigid-body transformations, not independent point displacements.
• Towards Next-Generation SLAM: A Survey on 3DGS-SLAM Focusing on Performance, Robustness, and Future Directions — 3D Gaussian Splatting offers a new paradigm for high-fidelity, efficient, and robust SLAM systems.
• OrbitForge: Text-to-3D Scene Generation via Reconstruction-Anchored Video Synthesis — OrbitForge converts text-generated videos into consistent 3D Gaussian Splatting scenes by anchoring reconstruction.
• Feed-Forward 3D Scene Modeling: A Problem-Driven Perspective — Feed-forward 3D reconstruction offers efficient, generalizable scene modeling by directly mapping 2D inputs to 3D representations.

See Counsel's argued verdicts on the open AI decisions leaders are weighing →

Open in AIssential →

Editorial summary, takeaway, and curation by AIssential. Original article published by cs.CV updates on arXiv.org.