CN104751463B  Based on the best way to choose a threedimensional model perspective sketch of the contour feature  Google Patents
Based on the best way to choose a threedimensional model perspective sketch of the contour feature Download PDFInfo
 Publication number
 CN104751463B CN104751463B CN201510145279.7A CN201510145279A CN104751463B CN 104751463 B CN104751463 B CN 104751463B CN 201510145279 A CN201510145279 A CN 201510145279A CN 104751463 B CN104751463 B CN 104751463B
 Authority
 CN
 China
 Prior art keywords
 profile
 dimensional model
 similarity
 model
 dimensional
 Prior art date
Links
Abstract
Description
一种基于草图轮廓特征的三维模型最佳视角选取方法 Based on the best way to choose a threedimensional model perspective sketch of the contour feature
技术领域 FIELD
[0001] 本发明涉及图像处理、计算机图形学领域，尤其涉及的是一种基于草图轮廓特征的三维模型最佳视角选取方法。 [0001] The present invention relates to image processing, computer graphics, and in particular relates to a threedimensional model based on the best view of the outline sketch feature selection method.
背景技术 Background technique
[0002] 近几年来，三维计算机图形学技术已经取得了长远的发展，成为了日常生活中不可或缺的一部分。 [0002] In recent years, threedimensional computer graphics technology has made longterm development, it has become an integral part of everyday life. 三维模型作为三维计算机图形学的基本要素尤其扮演着越来越为重要的角色。 Threedimensional model as the basic element of a threedimensional computer graphics, especially playing an increasingly important role. 为了在真实应用程序中取得更好的运行结果，就要求各类三维模型相关的分析建模算法拥有更高的计算精度。 In order to achieve better operating results in a real application would require the analysis of various types of threedimensional model modeling algorithm associated with a higher accuracy. 为三维模型自动地选取最佳视角是其中最为重要的算法之一， 常常作为其它三维模型相关算法的预处理工作。 Automatically select an optimal viewing angle for the threedimensional model is one of the most important algorithms, often as a threedimensional model preprocessing other related algorithms.
[0003] 三维模型最佳视角选取算法已经被广泛地运用于各类三维计算机图形学应用中， 其中包括:虚拟现实、三维模型检索、计算机辅助设计(CAD)以及三维多媒体等其它领域。 [0003] the best view of the threedimensional model selection algorithm has been widely used in various types of threedimensional computer graphics applications, including: Other virtual reality, threedimensional model retrieval, computeraided design (CAD) and threedimensional multimedia. 所谓三维模型最佳视角选取算法，即给定任意一个三维模型，为该模型计算出给定个数的观察视角并且使得这些视角最为符合人类的视觉感受。 The socalled threedimensional model to select the best view of the algorithm, i.e., given any threedimensional model, a model is calculated for the given number of viewing angle and viewing angle such that the most consistent with human vision.
[0004] 在现阶段的研究工作中，已提出了多种不同的三维模型最佳视角选取算法。 [0004] In the present stage of research, it has been proposed many different best view of the threedimensional model selection algorithm. 其中许多算法致力于探索三维模型的几何特征，例如构成三维模型顶点与面片之间的结构关系，同人类视觉系统之间的联系，这类算法包括模型显著性(Mesh Saliency)以及模型视角的熵(Viewpoint Entropy)等等。 Many algorithms devoted to the exploration of the threedimensional model geometric features, for example, a structural relationship between the threedimensional model vertices and patches, links between the human visual system, such algorithms include a significant model (Mesh Saliency) and a model perspective entropy (Viewpoint entropy) and so on. 它们的目标是通过分析三维模型的哪一部分最能引起人类的观察兴趣，以此来解决三维模型最佳视角选取的问题。 Their goal is to analyze the most part of the threedimensional model which can cause human observation of interest, in order to solve the problem the best view of the threedimensional model selected. 然而，对这一问题进行建模是非常困难的，这是因为要对三维模型进行精确的结构分析本身就是一个非常具有挑战性的工作。 However, the modeling of this problem is very difficult, because the threedimensional model to be accurate structural analysis itself is a very challenging task.
发明内容 SUMMARY
[0005] 本发明的目的在于提供一种基于草图轮廓特征的三维模型最佳视角选取方法，旨在从相关手绘草图中学习出能够反映人类视觉系统观察物体习惯的信息并以此计算出对应三维模型的最佳视角。 [0005] The object of the present invention is to provide an optimal viewing angle of the threedimensional model sketch outline feature selection method based on, we aim to learn from the associated handdrawn sketch to reflect the human visual system observation information object habits and thus calculate the corresponding threedimensional the best view of the model.
[0006] 本发明的技术方案如下： [0006] aspect of the present invention is as follows:
[0007] 一种基于草图轮廓特征的三维模型最佳视角选取方法，其包括以下具体步骤： [0007] A preferred threedimensional model perspective sketch profile based feature selection method, which comprises the following steps:
[0008] 步骤A:基于轮廓线条上下文环境的特征匹配算法，来计算草图和三维模型视角投影图的相似度，以此将所有给定的手绘草图映射到对应三维模型的视角上； [0008] Step A: featurebased matching algorithm context contour lines, and calculates the similarity dimensional sketch perspective projection model, in order to handdrawn sketches given all mapped onto the corresponding angle of view threedimensional model;
[0009] 步骤B:根据度量出的草图与三维模型视角的相似度，获取三维模型视角被手绘草图的映射概率，基于该映射概率设定约束条件，选取模型潜在最佳视角的正负样本数据库训练集； [0009] Step B: The similarity measure the threedimensional sketch of perspective model, obtaining a threedimensional perspective of the probability models are mapped freehand sketches, the constraint condition is set based on the probability map, select the best view of potential negative sample model database Training set;
[0010] 步骤C:利用词袋模型为每个三维模型构建特征向量，并基于正负样本，利用支持向量机训练出一个三维模型潜在最佳视角的分类器； [0010] Step C: bag of words using a threedimensional model constructing models for each feature vector, based on positive and negative samples, SVM training the potential of a threedimensional model of the best viewing angle classifier;
[0011] 步骤D:将三维模型视角的多样性引入到视角排序算法中，为每个三维模型选取出前N个给定个数的最佳视角。 [0011] Step D: introducing diversity into a threedimensional model Perspective View of the sorting algorithm, select the best view of the first N number given for each threedimensional model.
[0012] 所述的三维模型最佳视角选取方法，其在直接比较草图和三维模型视角的相似度之前，还要进一步进行如下操作:首先，将三维模型视角的投影图转化成与手绘草图相似的轮廓图；再，将所有轮廓图上轮廓线条内的像素点组成轮廓组，并将轮廓组根据相关度进行合并;然后，比较两轮廓组之间的相似度，最后根据轮廓组的相似度计算轮廓图的相似度。 [0012] The optimum angle of view threedimensional model of the selection method, prior to its similarity in threedimensional perspective sketches and direct comparison of the model, but also to further proceed as follows: First, the threedimensional perspective projection conversion model similar to handdrawn sketches FIG contour; again, the pixels in all the contour lines composed of the contour profile group, and the group profile in accordance combined correlation; then, the similarity between the two comparison groups contour, and finally the profile according to the similarity group calculating the similarity of the profile of FIG.
[0013] 所述的三维模型最佳视角选取方法，其每两个轮廓组gdPgj之间的相关度比较公式为： [0013] The method of selecting the best view of the threedimensional model, which compares each of the correlation between the two profile groups gdPgj the formula:
[0015] 其中，gi定义为初始轮廓组，Xi为该轮廓组在轮廓图上的平均位置，0i为该轮廓组的平均边缘方向，gj定义为对比轮廓组，Xj为该轮廓组在轮廓图上的平均位置，Θ」为该轮廓组的平均边缘方向，Gij为Xi和Xj的夹角大小。 [0015] wherein, gi is defined as the initial contour group, the average position for the Xi group on the contour of the profile, that average edge direction of the profile 0i group, GJ is defined as a contour comparison group, Xj of that group profile contour on the average position, Θ 'for the group average edge direction of the profile, Gij is the angle size of Xi and Xj.
[0016] 所述的三维模型最佳视角选取方法，其两轮廓组gdPgj之间形状的相似度为： [0016] The method of selecting the best view of the threedimensional model, the similarity between the shape of the two profile groups gdPgj of:
[0018]其中，θχ是轮廓组gx的边缘方向，dspa (gi，gj)是两轮廓组归一化后在各自轮廓图中平均位置的欧式距离，且的值为定值。 [0018] wherein, θχ edge direction of the profile groups of gx, dspa (gi, gj) is a set of two profile normalized in the respective average position of the rear profile of a Euclidean distance, and the setting value.
[0019]所述的三维模型最佳视角选取方法，其两轮廓组gdPgj之间的上下文信息加入相似度计算中，具体方法为:首先为每一幅轮廓图构建一个无向图。 [0019] The optimum angle of view threedimensional model of the selection method, which context information between the two groups gdPgj added contour similarity calculation, the specific method: First, a construct an undirected graph for each profile. 接着利用无向图计算轮廓组中任意两条路径的相似度;最后根据任意两条路径的相似度计算轮廓组的上下文相似度，具体的：为每一幅轮廓图构建一个无向图G= (V，E)，V为该无向图中的结点集合，E为图中的边集合，并定义 Next contours using no similarity to any group of two paths to FIG calculation; contextual similarity calculating final profile group based on the similarity of any two paths, specific: no construct a directed graph G for each profile a = (V, E), V for the drawing undirected nodes, E is the set of edges in the graph, and defining
为从结点即为轮廓组gl上出发的一条长度为η的路径，任意两条路径的相似度为： Is the length of a contour group gl starting from a node to a path η similarity of any two paths:
[0021] 其中 [0021] in which
是在路径： It is the path:
上的第k个结点，两轮廓组的上下文相似度为： Kth node on the two sets of context similarity profile:
[0023] 其中，P1"是所有从结点gl出发的长度为η的有序路径的集合，I P1nI表示集合P1"中所含路径的个数。 [0023] wherein, P1 "is all lengths starting from the node gl ordered set of path η, I P1nI denotes the set P1" contained in the number of paths.
[0024] 所述的三维模型最佳视角选取方法，其根据两轮廓组的相似度进一步获取两幅轮廓图Ci和Cj之间的上下文信息相似度，具体为： [0024] The method of selecting the best view of the threedimensional model, which similarity of two profile groups further obtains the similarity between the two context information Ci and Cj in accordance with a profile view, in particular:
[0026] 其中，gf表示在轮廓图c冲所包含的轮廓组，而I C11表示在c冲所包含轮廓组的个数。 [0026] wherein, GF represents a group profile contour punch included c, and c indicates the number of I C11 contour punch included in the group.
[0027]所述的三维模型最佳视角选取算法，其根据两轮廓组gdPm之间形状的相似度，并基于处于轮廓拐角处的关键点进一步获取轮廓图上下文信息相似度，具体为： [0027] The optimum angle of view of the threedimensional model selection algorithm, based on the similarity between the shape of the two profile groups gdPm, and further obtain the context information profile based on the similarity in the profile of the key points at the corners, specifically:
[0029] 其中，Kx为仅包含关键点的轮廓组集合，所有包含在集合Kx中的Pin为该轮廓图&的关键上下文信息描述算子。 [0029] wherein, Kx is the profile point group contains only key set, all included in the set Kx Pin of that profile & amp; key contextual information description operator.
[0030] 所述的三维模型最佳视角选取算法，其获取一副草图< 是从三维模型的视角\于被绘制出来的映射概率的具体方法为:对于每1 [0030] The optimum angle of view of the threedimensional model selection algorithm, a specific method for acquiring a probability map sketch <been drawn from the perspective of threedimensional model \ to: for each 1
[0032] 其中 [0032] in which
表示从三维模型视角· Threedimensional representation from the perspective of the model ·
计算所得的轮廓图。 The resulting profile is calculated.
[0033] 所述的三维模型最佳视角选取算法，其根据映射概率获取正负样本的方法为：当 [0033] The optimum angle of view of the threedimensional model selection algorithm, the probability of obtaining the mapping method based on positive and negative samples are: when
，将草图4映射到三维模型的视角 , 4 sketch map of threedimensional model to the angle
上，并且把所有满足此约束条件的三维模型视角作为训练时的正样本;计算对于草图集合S1I的所有 On, and the threedimensional angle of view satisfies all the constraints of this model during training as positive samples; sketch for calculating the set of all S1I
：的平均值，当三维模型视角t : An average value, when the threedimensional model viewing angle t
的平均值小于一个固定阈值时，把此视角作为负样本，整个采样策略的决策函数如下： When the average is less than a fixed threshold value, the angle of view this as a negative sample, the entire sample policy decision function as follows:
[0035] 其中 [0035] in which
：作为正样本，取〇则表示将其作为负样本，ξ为设定阈值。 : As positive samples, taken as the square indicates the negative samples, the set threshold [xi].
[0036] 所述的三维模型最佳视角选取算法，其在对三维模型视角进行排序时使用评价函数ti: [0036] The optimum angle of view of the threedimensional model selection algorithm that uses an evaluation function at the threedimensional perspective ti ranking models:
[0038] 其中，Φ (vi)是一个惩罚函数，α (·)是一个单调递减的函数。 [0038] where, Φ (vi) is a penalty function, α (·) is a monotonically decreasing function.
[0039]本发明的有益效果:本发明揭示了一种新颖的先验知识，认为当人们经常从某个视角去绘制三维模型时，则该视角是此三维模型的潜在最佳视角之一。 [0039] the beneficial effects of the invention: The present invention discloses a novel prior knowledge that often when people from a certain perspective to draw threedimensional model, the perspective is potentially one of the best views of this threedimensional model. 在将草图映射到三维模型视角的过程中，利用草图轮廓的上下文信息来进行相似度度量，有效地克服了草图含有大量形变噪音的问题。 During threedimensional perspective sketch map of the model, using the context information to the profile sketch similarity measure, effectively overcome the problem of deformation sketch containing a large amount of noise. 同其它三维模型最佳视角选取方法相比，本发明基于机器学习的方式拥有更稳定的性能，同时通用与不同种类的三维模型;选取的三维模型视角更符合人类视觉系统的直观感受，并且尤其适用于三维模型检索任务。 Compared with other methods to select the best angle of view threedimensional model, based on the embodiment of the present invention has a more stable machine learning performance, while common to different kinds of threedimensional model; model threedimensional perspective counterintuitive more selected human visual system, and in particular suitable for 3D model retrieval tasks.
附图说明 BRIEF DESCRIPTION
[0040] 图1是本发明整个框架的工作流程示意图。 [0040] FIG. 1 is a schematic flow diagram of the entire frame work of the present invention.
[0041] 图2&、213、2(：、2(1、26、2€是将轮廓图分割为轮廓组的示意图。 [0041] FIG. 2 & amp;, 213,2 (:, 2 (1,26,2 € is a schematic profile view of the profile is divided into groups.
[0042] 图3a、3b、3c是在轮廓图中提取关键轮廓组的示意图。 [0042] FIG. 3a, 3b, 3c is a schematic outline of the key group in the contour extraction in FIG.
[0043] 图4a、4b、4c是使用IoU来比较两个三维模型视角相似度的示意图。 [0043] FIG. 4a, 4b, 4c is a schematic view of the use of twodimensional perspective IoU models to compare similarity.
[0044] 图5是使用基于关键点的轮廓上下文信息相似度将手绘草图映射到对应三维模型视角上的示意图。 [0044] FIG. 5 is a similarity of context information based key contour points on the handdrawn sketches schematic threedimensional model corresponding to the perspective map.
[0045] 图6各类三维模型最佳视角选取算法在同样模型上的计算结果。 [0045] 6 perspectives of various threedimensional model of FIG best selection algorithm calculation results on the same model.
[0046] 图7是本发明提供的方法流程图。 [0046] FIG. 7 is a flowchart of a method of the present invention.
具体实施方式 Detailed ways
[0047] 为使本发明的目的、技术方案及优点更加清楚、明确，以下参照附图并举实施例对本发明进一步详细说明。 [0047] To make the objectives, technical solutions and advantages of the present invention will become more apparent, clear embodiment of the present invention is described in more detail below with reference to the accompanying drawings.
[0048] 本发明最主要的目的就是从相关手绘草图中学习出能够反映人类视觉系统观察物体习惯的信息。 [0048] The main object of the present invention is to learn from the relevant freehand sketches out the human visual system can reflect the information of the object observed habits. 首先，给定一个三维模型m与其对应的一系列手绘草图，本发明的目标是从这些数据中基于轮廓的上下文信息通过支持向量机训练出一个三维模型最佳视角分类器。 First, a series of handdrawn sketches given a threedimensional model corresponding to m, the object of the present invention is based on these data from the outline of the context information to train a threedimensional model by the optimum viewing angle classifiers SVM. 通过这个分类器，可以从三维模型ΠΗ的均匀包围球面上的三维视角空间V1中选取出该三维模型的若干个最佳视角，并且这些视角反映出了人类在手绘该物体时倾向于选择的观察位置。 With this classification, a viewing angle can be selected threedimensional space V1 surround the spherical surface of uniform threedimensional model ΠΗ withdrawn several optimal viewing angle of the threedimensional model, and these reflect the perspective of human observation of the object when the handdrawn prefer the position.
[0049] 参见图7,本发明提供的基于草图轮廓特征的三维模型最佳视角选取方法主要包含以下四个步骤： [0049] Referring to Figure 7, the present invention provides a threedimensional model of the best view of the outline sketch based feature selection method mainly includes the following four steps:
[0050] 步骤A:基于轮廓线条上下文环境的特征匹配算法，来计算草图和三维模型视角投影图的相似度，以此将所有给定的手绘草图映射到对应三维模型的视角上； [0050] Step A: featurebased matching algorithm context contour lines, and calculates the similarity dimensional sketch perspective projection model, in order to handdrawn sketches given all mapped onto the corresponding angle of view threedimensional model;
[0051] 步骤B:根据度量出的草图与三维模型视角的相似度，获取三维模型视角被手绘草图的映射概率，基于该映射概率设定约束条件，选取模型潜在最佳视角的正负样本数据库训练集； [0051] Step B: The similarity measure the threedimensional sketch of perspective model, obtaining a threedimensional perspective of the probability models are mapped freehand sketches, the constraint condition is set based on the probability map, select the best view of potential negative sample model database Training set;
[0052] 步骤C:利用词袋模型为每个三维模型构建特征向量，并基于正负样本，利用支持向量机训练出一个三维模型潜在最佳视角的分类器； [0052] Step C: bag of words using a threedimensional model constructing models for each feature vector, based on positive and negative samples, SVM training the potential of a threedimensional model of the best viewing angle classifier;
[0053] 步骤D:将三维模型视角的多样性引入到视角排序算法中，为每个三维模型选取出前N个给定个数的最佳视角。 [0053] Step D: introducing diversity into a threedimensional model Perspective View of the sorting algorithm, select the best view of the first N number given for each threedimensional model.
[0054] 上述方法步骤的可参见图1，图中展示了本发明的工作流程示意图。 [0054] The above process steps may Referring to Figure 1, there is shown a schematic diagram of the work flow of the present invention. 接下来，本说明书同样分成这几部分对本方法加以详细说明。 Next, the present specification likewise be divided into these portions of the detailed description of the present method.
[0055] 方法步骤A的目标是使用轮廓的上下文信息来度量草图和三维模型视角的投影轮廓图。 [0055] Step A target is to use the context information to measure the contour of a projection profile and the threedimensional perspective sketch model. 在直接比较草图和三维模型视角的相似度之前，还要进一步进行如下操作： Before direct comparison and threedimensional perspective sketch similarity model, but also to further proceed as follows:
[0056] 首先，将三维模型视角的投影图转化成与手绘草图相似的轮廓图。 [0056] First, threedimensional perspective projection model is converted into profile similar to the handdrawn sketches.
[0057] 再，将所有轮廓图上轮廓线条内的像素点组成轮廓线条组，简称“轮廓组”，并将轮廓组根据相关度进行合并。 [0057] Again, the pixel points in the contour plot contour lines on all contour lines composed of the group, referred to as "contour group", and contour are combined according to the correlation set.
[0058] 然后，比较两轮廓组之间的相似度，最后根据轮廓组的相似度计算轮廓图的相似度。 [0058] Then, the similarity between the two comparison groups contour, and finally the degree of similarity is calculated based on the similarity profile contour group. 具体解释如下： Details are as follows:
[0059] 本发明利用三维模型在每个视角形成的外轮廓线、闭合曲线、启发式轮廓线和模型边界线来产生最终的轮廓投影图。 [0059] The present invention utilizes the outer contour line in the threedimensional model is formed of each view, the closed curve, and contour heuristic model to generate a final boundary contour projection. 如图2 (a)和(b)所示，展示了使用此方法计算出的轮廓投影图示例。 As shown in FIG 2 (a) and (b), the projection view showing an example outline calculated using this method. 具体方法为:给定任意一幅轮廓图，首先在该图上使用边缘稀疏运算将草图上的轮廓线条变为一个像素点宽，如图2 (c)所示。 The specific method is: given an arbitrary profile, first used in the operation of FIG sparse edge contour lines on the sketch into a pixel width, as shown in FIG 2 (c) shown in FIG. 然后，使用Sobel算子计算每个像素点上的梯度方向。 Then, the subcalculate the gradient direction at each pixel using the Sobel operator. 通过上述操作结果会产生一幅非常稀疏的线条轮廓图，并且每个线条只有一个像素点宽度，每个像素点P都有一个边缘方向9[)度。 By the above operation will produce a result very sparse FIG contour lines, and each line width of only one pixel, each pixel P has an edge direction 9 [) degrees. 这一过程可以显著地减少初始轮廓图中所包含的大量噪音，同时拥有很高计算效率。 This process can significantly reduce a large amount of noise included in the initial profile, but has a very high computational efficiency.
[0060] 组成轮廓组的方法为:在预处理后的轮廓图的每个线条上均匀地采样一些像素点作为种子，基于这些像素种子，使用贪婪算法迭代地将种子周围八个方向上的邻居像素点合并起来直到它们的边缘方向夹角和大于一个阈值，本发明实施例中所述阈值为90度，这样就产生了该轮廓图上最初的轮廓组。 [0060] The method of the group consisting of contour: uniformly sampled points on the number of pixels per line in the preprocessed outline image as a seed pixel based on these seeds, greedy algorithm iteratively the eight directions surrounding neighbors seed until they are merged pixel edge direction angle is greater than a threshold value, embodiments of the present invention the threshold value is 90 degrees, thus creating a profile on the first contour group. 为每一个初始轮廓组gi定义，Xi为该轮廓组在轮廓图上的平均位置，Gi为该轮廓组的平均边缘方向，对比轮廓组gj定义，Xj为该轮廓组在轮廓图上的平均位置，θ」为该轮廓组的平均边缘方向。 Initial contour for each defined group gi, Xi for group on the average position of the contour of the profile view, Gi direction of the profile for the group average edge contrast profile defined gj group, Xj of that group, the average position of the profile on the profile the average edge direction θ "for the group's profile. 由此，每两个轮廓组之间的相关度则可以被定义为： Thus, the correlation between each of the two profile sets can be defined as:
[0062] 其中，Qij为Xi和χ」的夹角大小。 [0062] where, Qij to Xi and χ "angle size. 使用该公式将一幅轮廓图分割成多个在结构意义上的轮廓组。 The formula using an outline of the structure into a plurality of sense contour group. 公式1被用来进一步合并轮廓组直到任意两两相邻的轮廓组之间的相关度小于给定的阈值，本发明实施例中相关度衡量阈值设置为0.8,当小于0.8时则不进行轮廓组合并。 Equation 1 is used to set up further combined profile correlation between any two adjacent contour set smaller than a given threshold value, a measure of affinity embodiment the threshold is set to 0.8 the present invention, when the contour is not less than 0.8 and combination. 直观地来看，此方法可以进一步合并夹角较小的轮廓组，如图2(d)所示是一个最终轮廓组的结果示意图。 Visual point of view, this method may be further combined to contour a smaller angle, as shown in FIG 2 (d) is a diagram showing the result of a final profile group shown in FIG. 该轮廓分组方法在计算上是很简单的，而且，用于轮廓上下文信息的计算时同样非常有效。 The contour grouping method is computationally very simple, and, when the context information is used to calculate contour also very effective.
[0063] 使用内容的上下文信息来度量物体不同部分的相似度已在许多的三维模型分析研究工作中被证明是非常有效的。 [0063] content using the context information to measure the similarity of the different parts of the object has proved to be very effective in a number of threedimensional model analysis work. 轮廓的上下文信息指的是给定的轮廓线条同其周围的轮廓是如何连接在一起的，该特征往往提供了丰富的相似度信息。 Context information refers to the contour of a given profile with contour lines around how connected together, which often provide a rich feature similarity information.
[0064] 在本发明中实施例中，如果当两个轮廓组本身的形状以及其上下文信息相同时， 则认定它们更为相似。 [0064] In the embodiment of the present invention, if the profile when the two groups form and its context information itself is the same, the more similar they are identified. 本发明使用图论模型来进一步解释轮廓组相似度计算方法:定义两轮廓组gdM之间形状的相似度dapp (gl，gj)为： Using graph theory to explain the outline of the present invention is further sets of similarity calculation method: defines the degree of similarity between two dapp group gdM contour shape (gl, gj) is:
[0066] 其中，θχ是轮廓组gx的边缘方向，dspa (gi，gj)是两轮廓组归一化后在各自轮廓图中平均位置的欧式距离，且〇spa的值为定值，本发明实施例中为0.2。 [0066] wherein, θχ edge direction of the profile groups of gx, dspa (gi, gj) is the Euclidean distance of the two average position normalization contour set in the respective profile, and the setpoint value 〇spa, the present invention Example 0.2. 从上式中可以看出两个轮廓组被认为是相似的前提是当且仅当它们所处的轮廓图中的位置是相近的且它们的边缘方向也相近。 As can be seen from the above equation sets are considered two contours similar proviso that if and only if they are located in the profile of FIG position and direction of the edge thereof is also similar similar.
[0067] 为了将轮廓组的上下文信息加入到相似度匹配算法中，具体包括以下步骤： [0067] In order to outline the context information is added to the group of similarity matching algorithm, comprises the steps of:
[0068] 首先为每一幅轮廓图构建一个无向图。 [0068] First, a construct an undirected graph for each profile.
[0069] 接着利用无向图计算轮廓组中任意两条路径的相似度。 [0069] Next contours group similarity to any of the two paths is calculated without using FIG.
[0070] 最后根据任意两条路径的相似度计算轮廓组的上下文相似度。 [0070] Finally contextual similarity calculated profile based on the similarity of any two groups of paths.
[0071]具体解释为:为每一幅轮廓图构建一个无向图G= (V，E)，V为该无向图中的结点集合，E为图中的边集合。 [0071] DETAILED construed as: a profile for each construct an undirected graph G = (V, E), V for the drawing undirected nodes, E is the set of edges in the graph. 在V中的每一个结点表示一个轮廓组，当任意两个轮廓组在轮廓图上是物理位置相邻的话则使用一条边eeE将它们连接起来。 Each node in V represents a group profile, when the profile of any two adjacent groups are physical locations using an edge eeE then connect them in profile. 接着，定义 Next, define
为从结点（即为轮廓组)gl上出发的一条长度为η的路径，则任意两条路径的相似度为： Gl is a length from the starting node (i.e. profile group) η path, the similarity of any two paths:
[0073] 其中： [0073] wherein:
：是在路径 : Is the path
上的第k个结点。 Kth node on. 可以看出，路径 As can be seen, paths
是一系列从结点81开始的有序结点序列，这一序列描述了结Ag1周围的局部上下文结构信息。 Is a series of starts from the node junction 81 ordered sequence, this sequence is described taking the configuration information Ag1 local context around. 因此，可以通过寻找轮廓组gjPgj之间所有的有序结点路径之间的最相似匹配对，并把这些匹配对的相似度之和作为gi和gj的上下文相似度，其具体计算公式为： Therefore, by looking for the most similar match, and these matching similarities between all of the ordered set of nodes between the contour and the path gjPgj gi and gj as the contextual similarity, the specific formula is:
[0075] 其中，P1"是所有从结点gl出发的长度为η的有序路径的集合，I P1nI表示集合P1"中所含路径的个数。 [0075] wherein, P1 "is all lengths starting from the node gl ordered set of path η, I P1nI denotes the set P1" contained in the number of paths. 由于Pin能够有效地描述轮廓组81的周围所有上下文信息，则称Pin是轮廓组81的上下文信息特征描述算子。 Because Pin can effectively describe the profile group 81 surrounding all the context information, the context information is called Pin 81 wherein the group of contour description operator. 显然，当η为0的时候，公式4则退化成了只比较轮廓组之间形状相似度的公式2。 Clearly, when η is 0, the formula 4 formula degenerates to compare only the shape similarity between the contour 2 groups. 通过由公式4定义的相似度比较方法，就能够直接地比较两幅轮廓图之间基于上下文信息的相似度了。 By similarity comparison method defined by equation 4, it can be directly between the two comparative profile view of the context information based on the similarity.
[0076] 给定任意两幅轮廓图，首先将它们使用上文所述的方法构建出对应的无向结构图，接着为其中一幅轮廓图中的所有轮廓组使用公式4在另一幅轮廓图中寻找最为匹配的轮廓组，把这些轮廓组之间的相似度之和作为这两幅轮廓图的相似度。 [0076] Given any two profile view, they are first constructed using the method described above to the corresponding nonstructural view, and then a group in which all the contours in the contour plot using Equation 4 in a separate profile FIG find that most closely matches the contour of the group, the degree of similarity between the contour of the group and as the similarity of these two profile. 因此，两幅轮廓图C1 和W之间的上下文信息相似度的计算公式为： Therefore, calculation of similarity between the two context information C1 and the Wprofile of the formula:
[0078] 其中，gf表示在轮廓图Ci中所包含的轮廓组，而I Ci I表示在Ci中所包含轮廓组的个数。 [0078] wherein, GF represents a group profile contour Ci contained, and I Ci I represents the number of the contour Ci group comprising. 显然，当η为0的时候，公式5就退化成了仅考虑轮廓组自身基于形状相似度的情形，形式如下： Clearly, when η is 0, the formula became 5 degenerates to contour itself to consider only the case of the shapebased similarity, the following form:
[0080] 在公式5给出的基于上下文信息的轮廓图相似度计算方法中，需要指定每条路径的最大长度η。 [0080] In Formula 5 gives the contour plot of the context information based on a similarity calculation method, it is necessary to specify the maximum length of each path η. 当η为0时，则相当于只使用轮廓组的形状相似度来比较轮廓图。 When η is 0, the similarity of a shape corresponding to the outline set only to compare the profile of FIG. 使用较大的η 值能描述更多的全局结构信息，而当使用较小的η值时所能包含的信息就相对越小但计算速度越快。 Using a larger value η can be described more global structure information, and the information used when a small value of η can be contained in relatively smaller but faster calculation. 通过实验发现当η的值为3到5之间时能够产生最为稳定的结果，因此在本发明中将η的值固定在4。 Experimentally found that when the value of η will produce the most stable results when between 3 and 5, in the present invention thus fixing the value of η at 4. 图2 (e)和(f)显示了当η为4时，由同一个轮廓组所产生的两条不同路径。 FIG. 2 (e) and (f) shows that when η is 4, two different paths from the same group generated contour.
[0081] 用于匹配给定轮廓图的相似度匹配算法是本发明的核心算法之一，这是因为在之后的学习采样和训练阶段都要频繁地使用该算法，因此匹配算法必须拥有高效的计算速度。 [0081] Similarity matching algorithm for matching the given profile is one of the core algorithm of the present invention, since the algorithm to be used frequently in the learning and training phase after the sampling, and therefore must have efficient matching algorithm computing speed. 然而，在公式5中对于每一个轮廓组匹配对都需要在一个很大的搜索空间中寻找其最优匹配解，这是非常繁琐的计算过程，因而需要对此进一步加速。 However, in Equation 5 for each profile group match of the need to find the optimal matching solution in a large search space, which is very tedious calculations, so this needs to be further accelerated.
[0082] 在一副轮廓图中，其最大部分的上下文信息往往都被包含在了轮廓线条的拐角处，而与此相比单一的直线段则几乎不包含任何有用的信息。 [0082] In a profile, which is usually the largest part of the context information are included in the contour lines of the corner, and this compared to a single straight line segment almost does not contain any useful information. 因此，在寻找轮廓组的最优匹配解时，并不需要在所有的轮廓组中进行搜索，相反只需要在处于轮廓拐角处的轮廓组进行搜索即可。 Therefore, when looking for the best solution to match the profile of the group, you do not need to search across all the group's profile, instead only need to search in the profile set in the corner of the profile. 因此本发明进一步提出了基于关键点的轮廓图上下文信息匹配方法，具体算法如下： Accordingly the present invention further provides a profile view of the context information based on the key point matching method, the specific algorithm is as follows:
[0083] 使用诸如多尺度高斯算子(Difference ofGaussian)、Hessian算子以及HarrisLaplace算子的检测器能够有效地计算出给定轮廓图的拐角点，这些点被定义为该轮廓图的关键点。 Key [0083] Using such a multiscale Gaussian operator (Difference ofGaussian), Hessian operator HarrisLaplace operator, and a detector can be efficiently computed corner points of a given profile, which profile is defined as the point of point. 因此，在本发明中基于关键点的轮廓图相似度计算方法为:对于每一幅轮廓图cx，首先使用Harri sLaplace算子计算出所有该图中的关键点；接着根据每个轮廓组所处的位置即可得到仅包含关键点的轮廓组集合Kx，图3显示了这一计算过程，其中图3(b)为使用HarrisLaplace算子计算得到的关键点，图3 (c)为进一步计算得到的关键轮廓组集合； 然后，基于关键点的轮廓图上下文信息相似度被定义为： Accordingly, in the present invention is based on the critical points in a profile view of the similarity calculation method: for each of a profile cx, first using Harri sLaplace operator keys of all the calculated figure; then set in accordance with each profile can be obtained in which the position of the contour group contains only a set of keys Kx, Figure 3 shows this calculation process, wherein FIG. 3 (b) is calculated using the HarrisLaplace operator key, FIG. 3 (c) the key group is set profile further calculated; then, based on profile information key contextual similarity is defined as:
[0085] 并且，定义所有包含在集合Kx中的P1"(公式4)为该轮廓图cx的关键上下文信息描述算子。 [0085] Further, all included in the set defined in the Kx P1 "(Formula 4) described for the key operator profile cx is the contextual information.
[0086] 在确定了轮廓图的相似度匹配算法后，进一步阐述三维模型最佳视角分类器的学习算法。 [0086] After determining the similarity of the profile matching algorithm, the learning algorithm is further illustrated best threedimensional model viewing angle classifier. 在描述学习算法前，使用如下方法选定分类器学习算法所使用的数据库训练集。 Before describing the learning algorithm, using the following method selected database training set a classifier learning algorithm used. 在训练过程中，需要给定一个三维模型集合M以及同每一个三维模型meM相关的手绘草图集合Si。 In the training process, we require a threedimensional model given a set of handdrawn sketches and M associated with each of a set of threedimensional model meM Si. 在实验中，普灵斯顿三维模型基准库(Princeton Shape Benchmark，简称PSB模型库） 以及其相对应的草图库（由论文“Sketchbased Shape Retrieval”提供)被用来作为数据集。 In the experiment, PuLing dimensional model reference library Princeton (Princeton Shape Benchmark, referred to as model repository PSB) and its corresponding oxalyl gallery (provided by the paper "Sketchbased Shape Retrieval") is used as the data sets. PSB模型库把整个库分割成了两部分各自包含907个不同种类的模型，分别作为训练和测试集。 PSB model library to the entire library is divided into two parts each containing 907 different types of models, were used as training and test sets. 对于每一种类的三维模型，其提供了相同种类的手绘草图集合。 For each type of threedimensional model, which provides a set of handdrawn sketches of the same kind. 集合中的每一幅草图都由闭合的手绘曲线以及闭合的轮廓组成，基本符合本发明的实验需求。 Each of a set of handdrawn sketch by a closed curve composed of closed contour and, in line with the needs of the experiment of the present invention. 本发明使用PSB 提供的训练集来训练三维模型最佳视角分类器，而使用测试集合来验证本发明的计算性能。 The present invention provides PSB training set to train the threedimensional model of the best viewing angle classifier, using a test set to verify the calculated performance of the present invention.
[0087] 同其它的机器学习问题一样，为了学习出一个有效的能用于作视角分类的分类器，需要提供训练所需的正负样本。 [0087] The same problems as other machine learning, in order to learn a classifier can be used as a valid classification perspective, positive and negative samples needed to provide the necessary training. 然而，在原始的PSB数据库中并没有标示同三维模型视角相关的信息，因此，本发明采用如下方法进行正负样本的采样。 However, in the original database and PSB not marked with a threedimensional perspective of the relevant model information, therefore, the present invention employs a method for sampling the positive and negative samples.
[0088] 使用上文提出的基于关键点的轮廓图上下文信息相似度匹配算法将所有草图映射到对应的三维模型视角上，并进一步利用此关系筛选出每个模型的最佳和最差视角各自作为正负样本。 [0088]based profile key points raised above context information similarity matching algorithm maps all sketch to threedimensional model corresponding to the angle of view, and further use of this relation screened perspective top and bottom of each respective model as positive and negative samples. 其方法为： The methods are:
[0089] 首先，对于每一个在数据集中的三维模型me M，在其包围球面上均匀地选取K个视角并且为每个视角 [0089] First, for each data set in the threedimensional model me M, at each view angle to select the K uniformly surrounds the sphere and Perspective
计算器轮廓图 Calculator profile
本发明优选的取K的值为300时产生最为稳定的结果。 Produced the most stable results of the present invention is preferably 300 K is taken.
[0090]然后，对于每个属于三维模型mi的草图 [0090] Then, for each of the threedimensional model belonging mi sketch
使用基于关键点的轮廓图上下文信息相似度度量方法计算每个 Key profile based context information for each method for calculating the similarity measure
M•的相似度，则一副草图$是从三维模型的视角IVf:被绘制出来的概率为:对于每个； M • similarity, it is from the perspective of a sketch $ IVf threedimensional model: the probability of being drawn as follows: for each;
[0092] 其中，Cf表示从三维模型视角 [0092] where, Cf represents a threedimensional model from the perspective of
计算所得的轮廓图。 The resulting profile is calculated. 显然， Obviously,
将草图s映射到三维模型的视角if:上，并且把所有满足此约束条件的三维模型视角作为训练时的正样本。 The sketch s  mapped into perspective if threedimensional model: on, and the threedimensional perspective of all models satisfy this constraint as a positive sample during training. 为了收集负样本数据，需要计算对于草图集合&上的所有的平均值，当三维模型视角的平均值小于一个固定阈值时，则把此视角作为负样本。 For negative samples collected data needs to be calculated for a set of sketches & amp; on the average of all, when the average value is smaller than the threedimensional model of a fixed viewing angle threshold value, then this perspective as negative samples. 该负样本的采样策略的一个直观解释是，如果人类几乎从不从某个视角来绘制三维模型时，则认为此视角是一个很差的观察视角。 An intuitive explanation of the negative samples sampling strategy is that if humans almost never to draw threedimensional model from a certain perspective, it is considered that this perspective is a poor viewing angle.
[0093] 最后，为每个PSB数据库中含有的三维模型都采用以上策略采集正负样本数据，对于任意三维模型视角if,整个采样策略可以概括为如下决策函数： [0093] Finally, a threedimensional model database contained in each PSB are positive and negative samples using the above data acquisition strategy, for any threedimensional model perspective if, the entire sampling strategy can be summarized as follows decision function:
[0095] 其中 [0095] in which
1作为正样本，取0则表示将其作为负样本，且通过实验将阈值ξ设置为〇.〇5。 1 as a positive sample, which was taken as a 0 indicates a negative sample, and the experiment is set to the threshold value ξ 〇.〇5. 为了节省采样阶段的计算开销时间，所有这些结果都需预先离线计算出来。 In order to save computational overhead time of the sampling stage, all of these results are to be precomputed offline. 在采集完所需正负样本数据库训练集后，就可按如下过程训练三维模型最佳视角分类器。 After collecting the desired positive and negative training set sample database, threedimensional model can be trained classifier is the best view of the following procedure. 具体方法为： Specific methods are:
[0096] 先使用词袋模型来为每个三维模型的视角计算特征向量。 [0096] feature vector is calculated for each angle of view threedimensional model used to model word bags. 在训练数据集中，为了覆盖足够多种类的特征描述子，需要分别随机地从正负样本中选出总计一百万个关键上下文信息描述子;然后，使用kmedoids聚类算法将这些描述子构建出一个上下文信息描述子词汇表。 In the training data set, in order to cover enough variety of feature descriptors require total were selected at random one million key contextual information descriptor from the positive and negative samples; Then, kmedoids clustering algorithm using these descriptors Construction of a context vocabulary information descriptor. 在聚类结果中，所有的聚类中心W= {Wl}形成了整个词汇表，^表示了第i个描述子词汇的特征向量，则每个三维模型的视角特征即可表示为其对应轮廓图在上下文信息描述子词汇表中每个特征词汇出现的频率。 In the clustering result, all the cluster centers W = {Wl} forms a whole vocabulary, ^ represents the ith subword description of a feature vector, the viewing angle characteristics of each threedimensional model can be represented by its corresponding contour FIG word occurrence frequency of each feature descriptor information in the context vocabulary. 词汇表的大小W直接影响了后续分类结果的精度， 是一个非常重要的参数，所以本发明采用论文“Sketchbased Shape Retrieval”所提出的参数优化框架来获得其最优值，最终将IWI的值固定在800。 The size of the vocabulary  W  of a direct impact on the accuracy of the classification results of followup, is a very important parameter, the present invention uses a parameter optimization framework "Sketchbased Shape Retrieval" presented papers to obtain the optimal value, eventually IWI value 800 is fixed.
[0097] 令..表示三维模型视角tf,由词袋模型计算出的特征向量，本发明的目标是学习 [0097] indicates a threedimensional perspective so model .. tf, calculated by the model feature vectors bag of words, object of the present invention is to learn
出一个评价函f An evaluation function f
.来为每一个候选的三维模型视角预测出人类从该视角手绘此三维模型的可能性。 . To predict the possibility of a human handdrawn from this threedimensional model of the threedimensional perspective views for each candidate model. 使用支持向量机训练一个分类器，具体评价函数公式为： Using a support vector machine classifier training, specific evaluation function formula is:
[0099]其中t和b分别是由训练所学得的相关和有偏系数。 [0099] wherein t and b are learned from the training of the relevant and biased coefficients. 由于该问题中的特征向量大部分都是稀疏的，而LIBLINEAR是一个为稀疏特征特别优化过的SVM工具库，故本发明使用LIBLINEAR来训练分类器。 As most of the problem in the feature vector are sparse, and a sparse LIBLINEAR is characterized particularly optimized SVM tool magazine, so that the present invention is used to train a classifier LIBLINEAR. 值得指出的是，为了平衡在训练过程中使用的正负样本数量，分别从预先计算的正负样本集中等量地选出五千样本进行分类器的训练。 It is worth noting that in order to balance the number of positive and negative samples used in the training process, the positive and negative samples were concentrated from precalculated equally elect five thousand samples for training the classifier.
[0100] 在每个三维模型m的所有候选视角^£ V都使用公式10的评价函数取得评分后，即可通过将所有评分从高到低进行排序后选取得分最高的前几个视角作为此三维模型的最佳视角。 After [0100] Equation 10 uses an evaluation function in all threedimensional model viewing angle for each candidate m ^ £ V after obtaining the score, by all the ratings can be sorted in descending Select the highest score as the first few Perspective this threedimensional model of the best viewing angle. 然而，由于每个视角^是均匀分布在三维模型的包围球面上的，位置相近的视角有着相似的投影轮廓图，因此它们往往是非常相近的。 However, since each view ^ is uniformly distributed on a spherical surface surrounding the threedimensional model, and a position close perspective projection has a similar profile, so they tend to be very similar. 如果简单地选取前N个得分最高的^作为最佳视角的话就会造成结果中的视角都集中在三维模型的某一侧面的局部区域中，这在实际运用中是无用的。 If you simply select the highest score of the first N ^ as it will result in the best view of the results in perspective a localized area concentrated in the side of the threedimensional model, which in practice is useless. 为了在本发明算法返回的结果中尽可能地包含所有该三维模型的所有不同的最佳视角，则在对视角进行排序的时候需要将它们之间的多样性特点考虑在内。 In order to return the results of the algorithm of the present invention contains all of the best view of all the different threedimensional model is possible, we need to consider the diversity of the inner therebetween when the perspective of the sort. 因此，下文对本发明所使用的排序算法进行详细的描述，该算法鼓励排名较高的那些三维模型视角尽可能地分布在三维模型的不同位置。 Thus, hereinafter to the sorting algorithm used in the present invention is described in detail, the ranking algorithm encourage higher dimensional perspective of those models as possible distributed over different locations in the threedimensional model.
[0101] 令^表示使用公式10为三维模型视角^计算所得的初始得分，本发明引入了一个新的评价函数t，其定义为： [0101] order of 10 ^ represents a threedimensional model using the equation ^ Perspective of the calculated initial score, the present invention introduces a new evaluation function of t, which is defined as:
[0103]其中φ (V1)是一个惩罚函数，起作用是为了在结果中压抑相似三维模型视角的得分；而α (·)是一个单调递减的函数，它被用了控制惩罚函数的惩罚强度。 [0103] wherein φ (V1) is a penalty function, a function similar to depression scores in results of threedimensional perspective model; punishment strength α (·) is a monotonically decreasing function, which is controlled by the penalty function . 发现，α (·)函数只要能够快速地递减到〇即可，所以α (·)并不需要特殊地选取，在本发明中 Found, α (·) function as long as it can quickly decrements to square, the α (·) does not require specially selected, in the present invention,
,〇的值为〇. 2。 , Square square value. 2. 接着，惩罚函数则可写成： Then, the penalty function can be written as:
[0105] 其中T是一系列视角的集合，集合中的每个三维模型视角的排名都高于给定的视角Vi<JoU (Intersection over Union)被用来度量两个三维模型的视角是否相似，IoU被定义为两视角投影面积的交集除以它们的并集。 [0105] where T is a set of perspective, perspective threedimensional model of each set are ranked above a given perspective Vi <JoU (Intersection over Union) is used to measure the threedimensional perspective of two models are similar, IoU Angle is defined as the intersection of two divided by the projected area of their union. 图5显示了一个使用IoU来度量两视角相似度的例子，在图中根据计算可得视角(a)与(b)之间的相似度为0.87,而视角(a)与（c)之间的相似度为0.43,这之间的大小关系符合人观察所得的结果。 Figure 5 shows an example of using IoU to measure the similarity of two views, in FIG calculated according to the viewing angle (a) similarity between the (b) is 0.87, and the angle of view (a) and (c) between the the similarity was 0.43, in line with the size relationship between the results obtained human observation. 显然，惩罚函数Φ (V1)在对模型视角进行排序时会将同排名靠前视角非常相似的候选视角进行得分惩罚，从而达到压抑相似视角的作用，通过惩罚函数使得新的评价函数U将视角的多样性考虑在内。 Obviously, the penalty function Φ (V1) will be the same ranking in the perspective of the model is very similar to the perspective of sort of perspective candidates scoring the punishment, so as to achieve a similar perspective of repression, by making the new penalty function of viewing angle evaluation function U diversity into account. 为了不失一般性，定义当81 = ^时，T为空集。 Without loss of generality, when ^ = is defined as 81, T is the empty set. 此外，在排序过程中，为了取得更为稳定的视角位置，还为每个确定的三维模型视角使用meanshift算法寻找其评分的局部最优值作为最终结果，这样可以有效地减少由于均匀取样所带来的误差。 Further, in the sorting process, in order to obtain a more stable position of the viewing angle, as well as threedimensional perspective of each model is determined using meanshift algorithm to find the optimal value of its local score as the final result, which can effectively reduce the uniformly sampled the error caused. 以下是本发明完整的三维模型视角排序算法： The following is a complete model of the present invention is a threedimensional perspective sorting algorithm:
[0106] 输入:每个三维模型视角Vi e V的初始评分Si [0106] Input: Perspective of each threedimensional model Vi initial rates of Si e V
[0107] 输出：集合T，其包含了前N个此三维模型的最佳视角 [0107] Output: set T, which contains the first N best viewing angle of the threedimensional model
[0109] 为了证明本发明提出的三维模型最佳视角选取算法的有效性，首先比较了本发明提出的各类轮廓图相似度比较方法的性能，证明了基于关键点的轮廓上下文信息匹配方式是最为有效的；其次，通过在三维模型检索任务中将本发明同其它先进的三维模型最佳视角选取算法相比较，证明了本发明的有效性。 [0109] In order to prove the best view of the proposed threedimensional model of the present invention the effectiveness of the algorithm selection, first compared the performance of our proposed profile similarity comparison method of the present invention, demonstrating the contour matching method based on context information is the key point most effective; Secondly, in the threedimensional model retrieval object of the present invention with other algorithms to select the best view of the advanced threedimensional model as compared to prove the effectiveness of the present invention. 在实验中，如上文所述，在PSB三维模型数据库中对本发明进行训练和验证。 In the experiment, as described above, training and validation of the present invention in the threedimensional model database PSB.
[0110] 比较给定两个轮廓图的相似度在本发明中是一个非常重要的算法，这是因为它极大地影响了训练分类器的精度以及最终的选取结果。 [0110] Given two comparative profile view of the similarity algorithm is a very important in the present invention, because it greatly affects the accuracy of the results and the final selection of training a classifier. 所以，比较提出的所有轮廓图相似度计算方法是很有必要的。 Therefore, all proposed comparative profile similarity calculation method is necessary. 首先，在测试数据集中随机地采样了一百个三维模型及其相关的一幅手绘草图数据。 First of all, the test data set random sampling of one hundred threedimensional model of a handdrawn sketches and its associated data. 同时，邀请了十位用户分别为这些三维模型手工标定其所有相关的手绘草图是从哪个视角绘制的，随后把这些由用户手工标定的数据作为判别标准。 At the same time, invited 10 users for each of these threedimensional model of manual calibration of all its associated handdrawn sketches from which perspective is drawn, then these calibration data manually by the user as a criterion. 接着，分别使用上文所述不同的轮廓图相似度匹配方式将草图以同样的方式映射到三维模型的视角上，定义每种相似度匹配算法的准确度为： Next, using different profile similarity matching manner described above in the same manner as the sketch map to the angle of view threedimensional model, the definition of the accuracy of each of the similarity matching algorithm is:
[0112] 其中： [0112] wherein:
表示第k个用户为该三维模型选定的视角，η是用户的个数，在公式13中使用上文所述的IoU计算方式来比较两个三维模型视角间的相似度。 Represents the kth user is selected for the threedimensional model viewing angle, [eta] is the number of users, in the formula 13, to compare the similarity between the two threedimensional perspective IoU calculated using models described above. 最后，将每个轮廓图相似度计算方法在所有这一百个三维模型上的映射平均准确度作为评价该方法的标准。 Finally, the outline of each of the similarity calculation method for mapping the average accuracy on all one hundred of the threedimensional model as a standard evaluation method. 表1中展示了比较的所有轮廓图相似度计算方法，其中包括基于轮廓形状的（公式6)、基于轮廓上下文信息的(公式5)以及基于关键点的轮廓上下文信息相似度计算方法(公式7)。 Table 1 shows a comparison of all profile similarity calculation method, including those based on the contour shape (Equation 6), the profile context information (Equation 5) and the contour is calculated based on similarity of context information based on keypoint (Equation 7 ).
[0113] 表1不同轮廓图相似度计算方法的准确度 [0113] Table 1 Effect of profile similarity calculation accuracy of
[0115] 从表1中可以发现，在比较轮廓图的相似度时将轮廓的上下文信息考虑在内能够显著地增加匹配的准确度，但是计算复杂度也同时被极大地提升了。 [0115] can be found from Table 1, when comparing the similarity of the profile of the inner contour of the context information can be considered significantly increase the accuracy of the match, but also the computational complexity is greatly improved. 在牺牲了少量匹配精确度的前提下，将关键点检测技术加入到匹配算法中，能够节省大量的计算开销。 At the expense of a small number of matching accuracy of the premise, the key point detecting technology into matching algorithm, can save a lot of computational overhead. 因此，使用基于关键点的轮廓上下文信息相似度匹配算法是平衡了计算精度和速度之后的最优方式。 Thus, the use of context information based on contour matching algorithm key similarity balanced manner after the optimal speed and calculation accuracy. 图5列举了使用分发明提出的基于关键点的轮廓上下文信息相似度匹配算法来将手绘草图映射到其对应三维模型视角的一些结果。 Figure 5 lists the use of the contour information based on the context of the proposed subcritical points invention similarity matching algorithm to map some of its handdrawn sketch to threedimensional angle of view corresponding to the model results.
[0116] 由于选择三维模型的观测视角是一个比较主观性的任务，因此评价一个自动三维模型最佳视角选取算法的性能并不能直接用其在一个固定数据库上的准确率作为指标。 [0116] Since the selected observation angle of view threedimensional model is a more subjective task, thus automatically evaluate the best view of a threedimensional model selection algorithm performance and accuracy which can not be directly fixed on a database as an index. 在本发明中，通过把三维模型最佳视角选取算法运用到三维模型检索任务中，通过比较最终的检索准确度来间接地比较三维模型最佳视角选取算法的性能。 In the present invention, applied to threedimensional model retrieval task in the best viewing angle by the threedimensional model selection algorithm, by comparing the final retrieval accuracy optimal viewing angle to indirectly compare the performance of threedimensional model selection algorithm. 依照论文“Sketchbased Shape Retrieval”所述的方法来设置三维模型检索任务，包括数据采样、参数设置以及使用GALIF作为模型特征提取算法。 The method according to the paper "Sketchbased Shape Retrieval" to set the threedimensional model retrieval tasks, including data sampling, parameter and using a model feature extraction algorithm GALIF. 查准率查全率曲线(PrecisionRecall Curve)的曲线包含面积(AUC)被用来作为评价检索结果性能的指标。 Precision  recall curve curve (PrecisionRecall Curve) comprising the area (AUC) was used as an evaluation index of the performance of the search result.
[0117] 在整个检索过程中，分别使用不同的三维模型最佳视角选取算法来为每个模型选取候选视角供作为后续三维模型检索任务的输入，这些方法包括均匀分布的视角(在每个三维模型的包围球面上均匀采样N个视角）、基于模型显著性的方法（由论文“Mesh Saliency”提出）、最佳视角分类器（由论文“Sketchbased Shape Retrieval”提出）、基于互联网图片的方法（由论文“Webimage driven best views of 3D shapes”提出）以及本发明所述方法。 [0117] Throughout the search process, using different optimum angle of view threedimensional model selection algorithm to select a candidate for the angle of view for each model as an input of the subsequent threedimensional model retrieval tasks, these methods comprise the perspective of uniformly distributed (in each threedimensional surrounding the sphere model uniform sampling of N perspective), modelbased methods significance ( "Mesh saliency" made by paper), the optimal viewing angle classifier ( "Sketchbased Shape Retrieval" presented by the papers), Internetbased images of the method ( "Webimage driven best views of 3D shapes" made by paper) and the method of the present invention. 对于每个方法，通过不停地调整其参数来得到不同个数的最佳视角，直到所有方法的AUC指标都近似于0.23，显然使用越少三维模型视角个数就能达到此标准的AUC检索性能指标的视角选取算法的性能越好，详细的结果如表2所示。 For each method, a different number to obtain the optimal viewing angle constantly by adjusting its parameters, AUC index until approximately 0.23 in all methods are, apparently threedimensional model using the less number of viewing angle can be achieved retrieves this standard AUC the better the performance of the selected viewing angle performance of the algorithm, the detailed results are shown in table 2.
[0118] 表2不同三维模型最佳视角选取算法在被用于三维模型检索时的性能 [0118] Table 2 The best view of the threedimensional model selection algorithm performance when used in a threedimensional model retrieval
[0120] 如表2所示，本发明所使用的方法在使用了最少视角个数的情况下达到了最为优秀的AUC指标，尤其需要指出的是，本方法比其余方法最少节约了近两倍的视角个数，比均匀分布的视角选取方法节约了六倍的视角个数，这就证明了本发明拥有极高的视角选取准确度。 [0120] As shown in Table 2, using the method of the present invention achieves the most excellent in the case of using the AUC index number of the minimum angle of view, in particular, should be noted that the method of the present process than the remaining saving nearly twice the minimum number of perspective, the number of saving perspective six times than that of a uniform distribution of the selection method, which proved that the present invention has a high accuracy of the selected viewing angle. 图6展示了使用不用三维模型最佳视角选取算法所取得的结果，从上自下依次为本发明、基于互联网图片的方法、最佳视角分类器以及基于模型显著性的方法。 Figure 6 shows the result of using the best angle of view threedimensional model selection algorithm not obtained, since the order from the present invention, a method based on the Internet images, the best viewing angle and a method based on the classification model is significant.
[0121] 本发明揭示了另一种新颖的先验知识，同样能够有效地选取三维模型的最佳视角。 [0121] The present invention discloses novel priori knowledge of another, equally possible to effectively select the best angle of view threedimensional model. 本发明认为当人们经常从某个视角去绘制三维模型时，则该视角是此三维模型的潜在最佳视角之一。 The inventors believe that when people are often drawn from a certain perspective to the threedimensional model, the perspective is potentially one of the best views of this threedimensional model. 随着基于草图的三维模型检索算法的快速发展，许多包含对应手绘草图的三维模型资源库都已经被建立了起来。 With the rapid development of retrieval algorithm based on threedimensional model sketches, many threedimensional model repository contains the corresponding handdrawn sketches have been established. 这些数据库将手绘草图和三维模型之间建立起了有意义的联系，得益于此，手绘草图被用来建模本发明所提出的算法。 These databases will establish a meaningful connection between the handdrawn sketches and threedimensional models, thanks to this, handdrawn sketches are used to model the algorithm proposed by the present invention. 然而，同传统的图像相比，手绘草图是一种更为特殊的人类视觉信息载体。 However, compared with the traditional image, handdrawn sketches it is a more specific human visual information carrier. 在一幅手绘草图中，往往只包含了纯黑色的轮廓线条而缺少了颜色信息。 In a handdrawn sketch, often they contain only pure black contour lines and lack of color information. 并且这些线条由于是人工手绘而成，所以充满了各式各样的形变和噪音，草图的这些特性为特征匹配以及相似性度量增加了很大的难度。 And since the lines are painted from artificial, so full of noise and a wide range of deformation, these properties are sketches and feature matching similarity measure increases very difficult. 显然，将一幅手绘草图准确地映射到对应三维模型的视角处是一个非常具有挑战性的问题。 Obviously, a handdrawn sketch map exactly to the angle corresponding to the threedimensional model is a very challenging problem. 为了解决这个问题，本发明提出了一种使用草图轮廓的上下文信息来进行相似度度量的方法。 To solve this problem, the present invention provides a method of using the context information to the profile sketch similarity measure. 所谓草图的轮廓，即为组成草图中诸如直线、曲线等一系列边缘像素点的集合;而草图轮廓的上下文信息，则指的是这些轮廓片段是如何同其周围的轮廓互相组合起来的，这些信息往往能够表示所绘草图有意义的子部分(如动物的尾巴或者腿的轮廓）。 The socalled contour sketch, is the set of edge pixels composed of a series of sketches, such as lines, curves, and the like; and context information sketch of a profile, the profile refers to how these fragments are combined with each other around the contour, these information is often depicted can represent subportions (e.g., the tail or leg of an animal profile) sketch meaningful. 需要指出的是，草图轮廓的上下文信息总是蕴含了该草图的丰富特性，且能够有效地被用来度量它们之间的相似度。 It should be noted that the draft outline of the context information always contains a sketch of the featurerich, and can effectively be used to measure the similarity between them. 通过实验结果证明，这种相似性度量方法能够显著减少由于草图形变这类噪音所带来的影响，拥有稳定的性能。 The experimental results show that this similarity measure can significantly reduce the impact of noise caused by such deformation sketch, with stable performance.
[0122] 此外，草图轮廓的上下文信息还包含了能够反映出人类视觉系统习惯的共通特征。 [0122] Further, the context information further comprises the sketch profile common characteristic that reflects the human visual system habits. 例如，人总偏爱在草图的底部来绘制动物或桌子的四条腿，而动物的尾巴总被绘制在草图的水平两侧处。 For example, people always prefer to draw a sketch at the bottom of the animal or four legs of the table, while the animal's tail is always plotted on the horizontal both sides of the sketch. 鉴于此特点，本发明还提出了一个基于机器学习的方法，其使用草图轮廓的上下文信息且能够用来学习出一个通用的三维模型最佳视角分类器，该分类器可以为不同种类的三维模型自动地选取最佳视角。 In view of this characteristic, the present invention also provides a machinelearningbased method using a sketch of a profile and context information can be used to study the threedimensional model of a generic classifier optimal viewing angle, which may be classified into different types of threedimensional model automatically select the best viewing angle. 实验证明，本发明同其它三维模型最佳视角选取方法相比拥有好的性能，并且尤其适用于三维模型检索任务。 Experiments show that the best view of the present invention compared with other threedimensional model selection method has good performance and is particularly suitable for threedimensional model retrieval tasks.
[0123] 最近，Liu等人在论文“Webimage driven best views of 3D shapes”也提出了一种新颖的三维模型最佳视角选取算法，它为该领域引入了另一种全新的解决问题的思路。 [0123] Recently, Liu et al paper "Webimage driven best views of 3D shapes" also presents a novel algorithm to select the best view of the threedimensional model, which introduces another new problem for the field ideas. 这篇论文的与众不同之处在于，它并不直接分析三维模型同人类视觉系统之间的联系， 而是利用了本身含有人类观察物体的视觉信息的媒介去估计三维模型的潜在最佳视角。 Unusual in this paper is that it does not directly analyze the links between the threedimensional model with the human visual system, but to use the media itself contains information on human visual observation of the object to estimate the potential best view of the threedimensional model . 现有互联网上的图像资源被用来对这一方法进行建模，这类图像反映了人类在摄影时如何观察物体的视觉信息。 The image of the existing resources on the Internet are used to model this approach, such images reflect how people look at an object in the photographic visual information. 实验证明，该方法在用于三维模型最佳视角的选取上取得了非常好的性能。 Experiments show that the method has achieved very good performance in the selection of the best viewing angle for the threedimensional model.
[0124] 但是本发明同Liu的方法相比主要有三个不同之处。 [0124] However, the present invention has three main differences compared to the method of Liu. 首先，本发明使用了另一种能够反映出人类视觉系统观察物体习惯的媒体来解决这一问题，即手绘草图；手绘草图表达了人类偏爱从哪些视角去绘制三维物体，同互联网图片相比，能更为直接地对三维模型最佳视角选取问题进行建模。 First, the invention uses an alternative that reflects the human visual system to observe the object of media habits to solve this problem, namely handdrawn sketches; handdrawn sketches express human preference perspective from which to draw threedimensional objects, compared with the Internet pictures, can more directly for the best view of the threedimensional model to model selection problem. 其次，由于图片和草图所包含的视觉信息有着本质的区别：图片往往拥有丰富的颜色、纹理特征，而草图只包含了单色的轮廓线条；因此，本发明使用了与Liu的方法完全不同的特征提取方式，提出了基于轮廓线条上下文环境的特征匹配算法来计算草图同三维模型视角投影图的相似度。 Secondly, since the pictures and sketches included in the visual information is essentially different: often the picture is rich in color, texture features, and contains only a sketch monochrome contour lines; therefore, the present invention uses a completely different approach and Liu's feature extraction method is proposed wherein the similarity matching algorithm based on the context of the contour lines calculated with a threedimensional perspective projection view sketch model. 其三，Liu的方法在选取视角时高度依赖于三维模型的所属类别信息，即在计算最佳视角前必须先指定三维模型的具体类别且无法为未知类别的三维模型进行视角选取;本发明对此问题进行了改进，提出了一种基于机器学习的通用视角选取方法，通过学习三维模型最佳视角的共性，来为未知类别的三维模型进行最佳视角选取。 Third, Liu approach in selecting highly dependent on the angle of view threedimensional model information Category, i.e. must specify specific categories of the three dimensional model prior to calculating the optimum viewing angle and viewing angle can not be selected for the threedimensional model of the unknown class; of the present invention this problem has been improved, presents a general perspective based on machine learning method in selecting, to choose the best view for the threedimensional model of the unknown category of best perspective on learning through common threedimensional model.
[0125] 应当理解的是，本发明的应用不限于上述的举例，对本领域普通技术人员来说，可以根据上述说明加以改进或变换，所有这些改进和变换都应属于本发明所附权利要求的保护范围。 [0125] It should be appreciated that the present invention is applied is not limited to the abovedescribed example, those of ordinary skill in the art, can be modified or converted according to the above description, all such modifications and variations shall fall within the appended claims of the invention protected range.
Claims (9)
Priority Applications (1)
Application Number  Priority Date  Filing Date  Title 

CN201510145279.7A CN104751463B (en)  20150331  20150331  Based on the best way to choose a threedimensional model perspective sketch of the contour feature 
Applications Claiming Priority (1)
Application Number  Priority Date  Filing Date  Title 

CN201510145279.7A CN104751463B (en)  20150331  20150331  Based on the best way to choose a threedimensional model perspective sketch of the contour feature 
Publications (2)
Publication Number  Publication Date 

CN104751463A CN104751463A (en)  20150701 
CN104751463B true CN104751463B (en)  20171013 
Family
ID=53591082
Family Applications (1)
Application Number  Title  Priority Date  Filing Date 

CN201510145279.7A CN104751463B (en)  20150331  20150331  Based on the best way to choose a threedimensional model perspective sketch of the contour feature 
Country Status (1)
Country  Link 

CN (1)  CN104751463B (en) 
Families Citing this family (1)
Publication number  Priority date  Publication date  Assignee  Title 

CN106980623A (en) *  20160118  20170725  华为技术有限公司  Data model determination method and apparatus 
Citations (3)
Publication number  Priority date  Publication date  Assignee  Title 

CN102163343A (en) *  20110411  20110824  西安交通大学  Threedimensional model optimal viewpoint automatic obtaining method based on internet image 
CN102831638A (en) *  20120706  20121219  南京大学  Threedimensional human body multigesture modeling method by adopting freehand sketches 
CN103295025A (en) *  20130503  20130911  南京大学  Automatic selecting method of threedimensional model optimal view 
Family Cites Families (1)
Publication number  Priority date  Publication date  Assignee  Title 

WO2012000511A1 (en) *  20100629  20120105  3Shape A/S  2d image arrangement 

2015
 20150331 CN CN201510145279.7A patent/CN104751463B/en active IP Right Grant
Patent Citations (3)
Publication number  Priority date  Publication date  Assignee  Title 

CN102163343A (en) *  20110411  20110824  西安交通大学  Threedimensional model optimal viewpoint automatic obtaining method based on internet image 
CN102831638A (en) *  20120706  20121219  南京大学  Threedimensional human body multigesture modeling method by adopting freehand sketches 
CN103295025A (en) *  20130503  20130911  南京大学  Automatic selecting method of threedimensional model optimal view 
Also Published As
Publication number  Publication date 

CN104751463A (en)  20150701 
Similar Documents
Publication  Publication Date  Title 

Chen et al.  Similarity learning on an explicit polynomial kernel feature map for person reidentification  
Ji et al.  Spectralspatial constraint hyperspectral image classification  
Kapoor et al.  Gaussian processes for object categorization  
Andreopoulos et al.  50 years of object recognition: Directions forward  
Cabral et al.  Matrix completion for multilabel image classification  
Liu et al.  Recognizing human actions using multiple features  
Unnikrishnan et al.  Toward objective evaluation of image segmentation algorithms  
JP4516957B2 (en)  The method for performing a search for threedimensional objects, systems and data structures  
Kae et al.  Augmenting CRFs with Boltzmann machine shape priors for image labeling  
Liu et al.  Crossview action recognition via view knowledge transfer  
CN101315663B (en)  Nature scene image classification method based on area dormant semantic characteristic  
Su et al.  Multiview convolutional neural networks for 3d shape recognition  
Prosser et al.  Person reidentification by support vector ranking.  
Perakis et al.  3D facial landmark detection under large yaw and expression variations  
Wang et al.  Viewbased discriminative probabilistic modeling for 3D object retrieval and recognition  
Su et al.  Learning a dense multiview representation for detection, viewpoint classification and synthesis of object categories  
Kapoor et al.  Active learning with gaussian processes for object categorization  
Wang et al.  Linear neighborhood propagation and its applications  
CN101510257B (en)  Human face similarity degree matching method and device  
Cheng et al.  Global contrast based salient region detection  
Gould et al.  Decomposing a scene into geometric and semantically consistent regions  
Asian et al.  An axisbased representation for recognition  
CN103207879B (en)  Image index generation method and apparatus  
Grauman et al.  The pyramid match kernel: Efficient learning with sets of features  
Xiang et al.  Objectnet3d: A large scale database for 3d object recognition 
Legal Events
Date  Code  Title  Description 

C06  Publication  
C10  Entry into substantive examination  
ASS  Succession or assignment of patent right 
Owner name: TONGJI UNIVERSITY Free format text: FORMER OWNER: LIANG SHUANG Effective date: 20150722 

C41  Transfer of patent application or patent right or utility model  
GR01 