- 學位論文
高效能三維模型比對系統
Enhanced 3D Model Retrieval System through Characteristic Views Descriptor
摘要
由於電腦圖學的突飛猛進，愈來愈多地方可以看到其應用，其中又以三維技術的運用最為顯著。工業方面，電影、遊戲等娛樂工業大量地使用三維技術，連帶地造成大量的三維工具雨後春筍般問世，舉凡例如三維模型建模、三維場景算圖、多媒體製作工具等。由於簡單易用的網路三維（Web3D）開發工具興起（Cult3D, Shockwave3D, Pulse 3D…），使得愈來愈多非專業人力投入相關創作。然而這些工具是為專業人員使用而設計的，且製作三維模型仍舊需要專業的美術背景。網路上已經有愈來愈多人提供免費供人下載的三維模型，因此一個簡單好用又有效率的三維模型搜尋系統就變得迫切需要了。 過去幾年雖有釵h研究團隊先後投入三維模型搜尋與比對之相關技術，囿於此課題較二維圖形的比對來得困難，且缺少公定之比對資料庫，使得各家數據無從比較。直到美國Princeton大學提出了一套三維模型資料庫（Princeton Shape Benchmark database），彙整了數個當代技術，做出了整合性的比較，其中由台灣大學陳鼎勻博士等人的「以視覺相似度為基礎的三維模型搜尋技術」的比對正確率最高。他們透過在一個三維模型周圍照了100張平行投影的影像，再將這100張影像參數化，使得兩個三維模型之比對得以轉換為兩群參數比對。此方法比對正確率雖高，但線上比對速度則明顯較其它方法慢了數十到一百倍左右。這也是本篇論文的主要研究課題。 因此，我們提出了一個大膽的假設：一個三維物體只需要藉由少數的幾張影像就能夠很準確地描述它的外形。對每一個三維模型，我們提出了Orthogonal Visual Hull這個方法，找出6張特徵影像（Characteristic Views）拿來以資比對，並且為了彌補使用較少張的影像所可能造成的比對率下降，具有深度資訊之影像也拿來使用，以彌補影像採用張數過少所引起的比對率下降。是以總計本方法採用了特徵影像上的輪廓、形狀、深度值、扁平化程度和圓的程度等資訊來做比對，分別藉由Fourier Descriptor、Zernike Moments Descriptor、Generic Fourier Descriptor、Eccentricity和Circularity等技術來參數化。 本篇論文的主要結論與貢獻如下：(1) 此方法透過僅用6張特徵影像來取代原來的100張影像的方式，使得線上的比對速度快了約莫50到100倍。(2) 加進深度值的資訊，使得比對正確率能保有甚或超過原陳鼎勻博士等人的結果。(3) 提出並建構了特徵影像和人類視覺認知的關聯性，並嘗試帶出三維模型比對與人類視覺認知在未來是個很前瞻性的研究課題。
並列摘要
With the distinct evolution in computer graphics, more 3D applications are developed, including modeling tools, rendering software and authoring applications. Those applications were ever designed for professionals, but now even the crowd can take advantage of them to do something favorably such that people start to demand more and more 3D models just like they search in Google for everyday information, homework and others. Through nice and easy-to-use modeling tools and more people get involved in creating 3D models, a large number of 3D models are available on Internet. The techniques for 3D model retrieval then become necessary. Much research of 3D model retrieval techniques has been proposed recently and the system based on LightField Descriptors by Chen et al. have been proven as the most effective system in retrieval accuracy until now except its high on-line computing cost compared to other techniques. In this thesis, an enhanced 3D model retrieval system based on Characteristic Views Descriptor is proposed to speed up on-line computation and achieve the retrieval accuracy of the system by Chen et al. To the former, instead of one hundred orthogonal projections of a 3D model, six projections as Characteristic Views are required to perform matching between 3D models, which eminently boost on-line matching speed about 50 to 100 times faster. To the latter, except the original image metrics, distinctive Generic Fourier Descriptor on range data is employed to recover the accurate loss caused by fewer projections. And the system achieves nearly the same retrieval accuracy as Chen et al.’s system. Moreover, a brand new idea for pose estimation by Orthogonal Visual Hull is also proposed. We conclude that the new system overmatches Chen et al.’s system in time, space and retrieval accuracy and thus the contribution of this thesis is to provide a more practical 3D model retrieval system to the public.