程序式合成的高階控制(I)

Abstract

程序式合成技術被廣泛的使用在產生材質圖案、地形模型、植物模型、動物毛皮 紋路、建築物以及城市模型等。這類技術在大規模合成的應用上很有效，而在資料儲 存空間有限或是資料不易獲得的情況下特別有用。在程序式材質合成法中，如柏林雜 訊、小波雜訊及Gabor 雜訊等早已被廣泛的應用在電腦遊戲、視覺特效及電腦動畫中。 而在近幾年，由於建構虛擬環境與自然及都市生態模擬的需求，程序式合成技術也被 應用在植物及建築物的建模生成上。 儘管程序式合成技術的應用不勝枚舉，程序式合成的高階控制卻是一個極少被注 意的問題。使用者在使用程序式合成方法時常常需要反覆調整很多的低階參數以達到 想要的結果，這個過程往往非常繁瑣耗時。事實上，目前在材質雜訊函數的使用上， 並不存在一個有效的控制方法。而在植物及建築物模型的程序式合成上，雖然目前已 經有ㄧ些研究在開發更友善更直覺的方法，例如以影像為基礎的建模或是描繪式介面 等，這些方法仍然著重在重建出與影像或繪圖中的物體在視覺上相似的三維模型。由 於這些方法所重建出的模型，不ㄧ定符合原有物體的物理或生物合理性，所以這些模 型也無法被應用在物理或生物模擬中，其應用受到很大侷限。 在這個計畫中，我們將發展一個程序式合成的高階控制統合架構。這個架構的核 心是一個基於最佳化與機器學習的方法，能夠依據使用者輸入的簡單繪圖、範例圖形 或影像以及其他高階指令等，自動選擇程序式合成系統的模型或法則並調整參數，以 產生使用者想要的合成結果。由於程序式合成的模型或法則都具有物理或生物合理 性，我們的方法所獲得的參數及生成結果都自然具有物理或生物合理性。我們亦將結 合不同輸入媒介，配合學習機制，發展更智慧的高階控制介面。 我們預期這個計畫的貢獻是：(1)提出一個程序式合成的高階控制的新穎觀念；(2) 研發出一個透過模型或法則自動選擇及參數最佳化，進行程序式合成高階控制的統合 架構。(3)由於我們所提出的高階控制統合架構，程序式合成將有更廣泛的應用。

Procedural synthesis techniques are widely used in generating texture patterns, terrain geometry, botanical models, animal coatings, architecture, and urban models. They are very efficient for large-scale synthesis and are particularly useful when the storage of data is limited or the acquisition of data is difficult. Procedural texture synthesis approaches, such as Perlin noise, wavelet noise, and Gabor noise, have been used in computer games, visual effects, and computer animation. In recent years, procedural synthesis techniques have also been applied to botanical and architectural modeling due to the needs in virtual environment creation and simulation. Despite numerous applications of procedural synthesis techniques, high-level control of procedural synthesis has not been addressed in the past. As a result, users often need to tune many low-level parameters iteratively to achieve desired results, which is tedious and unintuitive. For example, it is very common to tune many parameters in Perlin noise to achieve desired appearance. In fact, there is no effective and efficient way to control the texture noise functions. For procedural synthesis of botanical and architectural models, although there have been research efforts on developing more user-friendly and intuitive approaches, such as image-based modeling or sketch-based interface, these approaches are still limited to reconstructing 3D models that are visually similar to the objects in images or sketches. Therefore, the constructed models may not be physically or biologically sound, which prevents them from being applied to physical or biological simulation. In this project, we will develop a unified framework for high-level control of procedural synthesis. The core of our framework is an optimization-and-learning-based approach that can automatically select procedural models/rules and adjust parameters used in a procedural synthesis system. As physically-sound models/rules are used in our procedural synthesis process, the acquired parameters and rules and synthesized results would be physically-sound. We will also investigate to develop a more intelligent interface that integrates multiple input modalities and adapts to a user by learning from the user’s corrections on the results. The potential contributions of this project will be: (1) introducing a novel concept for high-level control of procedural synthesis; (2) proposing a unified framework to intuitively control procedural synthesis approaches through automatic model/rule selection and parameter optimization; (3) extending applications of procedural synthesis due to the proposed framework.