四維高斯潑灑資料擷取系統建構與測試

Abstract

隨著高斯潑灑等神經渲染技術的興起，三維重建進入了低成本、高品質的新時代。然而，在進行四維場景重建時，為了維持重建品質，仍需仰賴多台高階攝影機與專攝影棚的拍攝環境，對一般使用者而言門檻仍高、成本不易負擔。本研究旨在探討如何在低成本、不穩定的場地條件下，建構一套可用於四維高斯潑灑訓練的高品質資料拍攝與處理流程。所提出的系統涵蓋了多機架設策略、影像同步、色彩校正、前景擷取與相機參數計算等環節，並以商用iPhone裝置實作驗證其可行性。 本研究採用V^3模型（Viewing Volumetric Videos on Mobiles via Streamable 2D Dynamic Gaussians）進行訓練與重建，驗證所建立資料集在不同實驗配置下的有效性。透過本方法，期望能降低四維重建技術的門檻，促進其在教育、娛樂與個人應用中的普及化。

Neural rendering techniques such as Gaussian Splatting has significantly lowered the cost of high-quality 3D reconstruction. However, its extension to 4D reconstruction still typically relies on multi-camera setups and controlled studio environments, which pose practical barriers for general users. This thesis proposes a low-cost and flexible data acquisition pipeline for 4D Gaussian Splatting using consumer-grade iPhone devices. The pipeline integrates synchronized multi-camera capture, color correction, foreground extraction, and camera parameter estimation. We integrate the captured data into a modified version of the V^3 framework (Viewing Volumetric Videos on Mobiles via Streamable 2D Dynamic Gaussians) to validate the feasibility and effectiveness of our approach. Experimental results demonstrate that high quality training data can be acquired under non-professional conditions, thereby making 4D reconstruction more accessible to a wider range of users.