以感知為基礎的骨架對應

Abstract

物體對應就是在兩個以上的物體元素之間建立有意義的關聯。 而實際應用中大多需要語義之間的關聯。 但是在只有空間和幾何資訊的條件下，要找到語義的對應是很困難的問題。 因為我們無法保證物體的幾何和語義之間具有一定的關聯性，而物體之間的姿勢和表面上的細節也會增加找到正確對應的難度。 然而人類卻不會受到這些問題的影響，因為我們熟悉物體的形狀、特徵以及功能，可以憑藉感知找到物體之間的對應。 因此我們提出一個新的對應方法，結合感知的概念模擬人類尋找對應的過程來解決這些問題。

首先，我們利用對稱性將骨架簡化成一個形式統一的圖，再來在此圖上分析其結構，將其分割為三個具有各自意義的主要部位。 其中這些主要部位之間的對應是固定的，可以在分割主要部位時確定對應關係。 在骨架的結構都分析且分割完成後，利用骨架的三個主要部位之間的對應關係和骨架的對稱性來計算整體的對應。 此對應的結果將會是唯一的，並且是我們認為的最佳解。 我們的方法可以接受骨架之間具有不同拓撲結構、姿勢和幾何性質，仍成功找出合適的對應結果。

Shape correspondence is about establishing meaningful relationship between two or more elements of shapes. In real world applications, semantic correspondence is heavily required. But with only spatial and geometric information, finding semantic correspondence is a difficult problem since we can't guarantee the correlation between the geometry and semantic of shape, with poses and surface details between shapes further increase the difficulty of finding correct correspondence. However, since human are familiar with the feature, functionality of shape components and appearance, we can find the correspondence between the shapes with perception without being troubled by problems above. Therefore, we propose a new method which simulates human's perception to find the correspondence.

We first simplify skeleton into a general graph format by using symmetry, and then, by analyzing the resulting graph, we can segment it into three identical main parts. Note the correspondence between these main parts is fixed when we analyze and segment the skeleton structure. After finding the main parts, we can use the mapping between parts in different models and skeleton symmetry to construct final correspondence. The constructed result is unique and expected to be an optimal solution. Our method can find appropriate correspondence result even if poses, topology structures, or geometry features are different between skeleton.