基於時空重採樣的遮擋問題分析與改進方法

Abstract

光線追蹤因為計算耗時，在應用上長期受到限制，以往大多只能用於離線算圖的場合，例如：電影、動畫等非即時互動的領域。自從 2020 年 ReSTIR（Reservoir-based Spatio-Temporal Importance Resampling）技術被提出來之後，大幅提升了即時光線追蹤的效率，使得光線追蹤在即時互動的應用上變為可能，推動了光線追蹤在遊戲及虛擬實境等領域的發展。然而，ReSTIR 在處理場景中被遮蔽後重新顯露（disocclusion）的區域時，仍面臨顯著挑戰。當視角改變或是場景發生變動，導致原本被遮蔽的區域突然顯露的情況下，時空重採樣找不到前一幀的有效樣本，容易因樣本不足而導致雜訊明顯增加，進而影響整體的畫面品質。本研究針對 ReSTIR 時空重採樣技術在 disocclusion 區域的問題進行深入分析，並提出一種基於迭代時空重採樣的方法，有效改善即時算圖中的 disocclusion 雜訊問題。本方法僅需極少的額外運算資源，並通過多個實驗場景進行驗證。實驗結果表明，與 ReSTIR GI 相比，本方法在多個場景下實現 RMSE 減少最高達 81.5%，平均減少約 58.5%，且平均時間開銷僅增加 2.68 ms，顯示本方法在 disocclusion 算圖品質提升與處理效能間的優異平衡，為 disocclusion 問題提供了有效的解決方案。期望本研究能進一步促進即時光線追蹤的應用發展，並為遊戲、虛擬實境等即時互動場景帶來更真實且流暢的視覺體驗。

Ray tracing has long been constrained in practical applications due to its computational cost, traditionally limited to offline rendering scenarios such as film and animation, which do not require real-time interaction. In 2020, the advent of ReSTIR (Reservoir-based Spatio-Temporal Importance Resampling) has greatly improved the efficiency of real-time rendering, enabling ray tracing in interactive applications and advancing its use in fields like gaming and virtual reality. However, ReSTIR still faces significant challenges in handling disocclusion regions in a scene. When the viewpoint or scene changes, causing previously occluded regions to become visible, spatiotemporal resampling often fails to find valid samples from the previous frame, leading to increased noise due to insufficient samples and degrading overall image quality.This study conducts an in-depth analysis of the disocclusion issues in ReSTIR's spatiotemporal resampling and proposes an iterative spatiotemporal resampling method to effectively mitigate disocclusion noise in real-time rendering. The proposed method requires minimal additional computational resources and has been validated across multiple experimental scenes. The results demonstrate that, compared to ReSTIR GI, this method achieves a reduction in RMSE of up to 81.5%, with an average reduction of approximately 58.5%, while incurring an average time overhead of only 2.68 ms, showcasing the method's excellent balance between enhancing disocclusion rendering quality and maintaining processing efficiency, and providing an effective solution to the disocclusion challenge. This highlights the method's effective balance between enhancing disocclusion rendering quality and maintaining computational efficiency, offering an effective solution to the disocclusion problem. It is hoped that this research will further promote the development of real-time ray tracing and contribute to more realisticand fluid visual experiences in gaming, virtual reality, and other interactive scenarios.