Year: 2026
Type: article
Abstract: Dynamic-scene 4D reconstruction technology has great application potential in film and television production, virtual reality, and smart cities. However, existing methods generally suffer from limitations such as constraints on ultra-dense or single-view input, end-to-end black-box design, frame-by-frame independent output, and high computational requirements. To address these issues, this paper proposes a modular, streaming, asynchronous, and spatiotemporally coherent 4D reconstruction method. Its main innovations are: (1) native support for sparse multi-view video input; (2) a modular architecture with four decoupled and independently replaceable modules; (3) a temporal-stream depth estimation that propagates and optimizes depth across consecutive frames; and (4) a hierarchical keyframe scheduling strategy for streaming asynchronous processing that achieves online / near-online processing. By using sparse multi-view video as input and a modular design—where each core stage is decoupled and independently replaceable—the method outputs spatiotemporally coherent dynamic points. Experiments on multiple datasets show that the proposed method can achieve online, incremental, dynamic 4D reconstruction of multi-view video data while ensuring reconstruction accuracy and completeness.
Source: Scientific Reports
Authors: Yang Dong, Jiaxuan Song, Dazhao Fan, Song Ji
Institution: PLA Information Engineering University
Language: English
Topic: Advanced Vision and Imaging
Subfield: Computer Vision and Pattern Recognition
Field: Computer Science
Domain: Physical Sciences
Open Access status: gold
Estimated APC paid: $2,690
OpenAlex ID: W7211903150
FWCI:
Cites:
Cited by:
Related to: