Date of Award

12-2025

Document Type

Thesis

Degree Name

Master of Science (MS)

Department

Electrical and Computer Engineering

Committee Chair/Advisor

Lan Zhang

Committee Member

Tao Wei

Committee Member

Luyang Zhao

Abstract

Dynamic 3D videos are now widely used in VR, AR, and telepresence, where users often want to change the style or appearance of objects such as clothing over a whole sequence. However, editing such dynamic 3D content is still hard, because appearance, geometry, and motion are tightly coupled, so even a simple color change must stay consistent in all views and at all time steps. Existing editing methods based on 4D Gaussian Splatting often work in a frame-by-frame way, which is slow and makes it difficult to keep temporal and multi-view consistency.

This thesis proposes Editable-4DGS, a representation-centric framework that separates appearance from motion on top of 4DGS. We first reconstruct a 4DGS scene and learn a canonical UV texture that stores the appearance of the 4D geometry. Then we use optical-flow–conditioned Gaussian flow to track each Gaussian over time and to compute a UV trajectory for each frame, while the canonical texture itself stays fixed. On top of this, per-Gaussian semantic features allow us to select meaningful regions, such as a shirt, and perform semantic-aware retexturing directly in UV space. In this way, editing a dynamic scene is reduced to changing a single canonical texture, and the system automatically propagates this change to all frames. Experiments on DyNeRF and Tensor4D datasets show that Editable-4DGS produces more temporally and multi-view consistent retexturing than the baseline, and also provides better semantic control.

Share

COinS
 
 

To view the content in your browser, please download Adobe Reader or, alternately,
you may Download the file to your hard drive.

NOTE: The latest versions of Adobe Reader do not support viewing PDF files within Firefox on Mac OS and if you are using a modern (Intel) Mac, there is no official plugin for viewing PDF files within the browser window.