低頻事前輻射轉移計算用於動畫模型之參數壓縮

Abstract

事前輻射轉移計算是種將全域照明實現在即時運算的應用軟體中的技術。轉移函式表示著光線複雜的轉移傳輸，透過事前運算，我們可以在程式執行時利用光線傳輸是線性的特質把對入射光的積分簡化 成簡單的線性運算。 對於可變形的3D模型，事前運算會產生大量的資料。基於轉移函式和動畫參數彼此間有著線性關係的假設，我們提出一個方法從一組參考模型中推演出兩者的相依性。在程式執行時，轉移函式中的每個參數可以由目前動畫參數以及每個頂點參數的內積來近似。這方法產生很微量的計算， 而且因為不論動畫的長度，每個頂點都只帶有一組參數，藉此我們可以節省儲存空間。 我們主要的貢獻在於提出一個容易在繪圖晶片上實作的方法，可以快速地近似出轉移函式。群聚主要分量分析是事情輻射轉移計算中普遍在使用的壓縮方法，以群聚主要分量分析為對照，我們藉由比較錯誤率和使用空間分析所提出的方法。因為動畫中的轉移函式是在執行時近似出的，我們不需要對轉移函式作費時的事前處理。

Precomputed radiance transfer (PRT) is a technique to bring global illumination into real-time applications. The method precomputes the complicated transport of light represented by a transfer function and exploits the linearity of light transport to reduce the light integral to a simple linear algebra operation. For a deformable model composed with a sequence of meshes, data from precomputation occupy plenty of storage. Based on the assumption that the transfer function is linear to the animation parameters, we represent a method to infer the linearity with a set of reference poses as its input. At run-time, each coefficient of transfer functions are approximated with dot product between the current animation parameters and per-vertex coefficients. The method introduces little computational overhead and saves storage by storing only one set of coefficients at each vertex regardless the length of an animation. Our main contribution is to propose a simple algorithm which can be easily implemented on GPU to approximate transfer functions. We analysis the method by comparing error rates and storage to CPCA, which used to be a compression strategy of PRT in a static scene. Since transfer functions of a given animation are approximated at run-time, theres is no costly precomputation on transfer functions required.