Bag-of-Feature-Graphs: A New Paradigm for Non-rigid Shape Retrieval T i n gbo HOU, X i a ohua HOU, M i n g ZHON G a n d Ho n g QI N De p a rtm ent o f Co m p u te r S c i enc e S to ny Bro o k U n i ve rsit y ( S U N Y S B) Department of Computer Science Center for Visual Computing ICPR 2012 Nonrigid Shape Retrieval Shape Query Shape Database Retrieved Shapes … … Department of Computer Science Center for Visual Computing ICPR 2012 Overview of BoFG Inspired by the ideas from Bag-of-Words (BoW) and SpatialSensitive Bag-of-Words (SS-BoW) Feature-driven Concise and fast to compute Spatially informative Department of Computer Science Center for Visual Computing ICPR 2012 Previous Works Relevant to This Project Bag-of-Words 1. 2. 3. 4. Y. Liu, H. Zha, and H. Qin. CVPR, 2006. H. Tabia, M. Daoudi, J. P. Vandeborre, and O. Colot. 3DOR, 2010. R. Toldo, U. Castellani, and A. Fusiello. VC, 2010. G. Lavoué. 3DOR, 2011. Shape Google (Spatially-Sensitive Bag-of-Words) 1. M. Ovsjanikov, A. M. Bronstein, L. J. Guibas and M. M. Bronstein. NORDIA, 2009. 2. (SI-HKS) M. M. Bronstein and I. Kokkinos. CVPR, 2010. 3. A. M. Bronstein, M. M. Bronstein, L. J. Guibas, and M. Ovsjanikov. ACM TOG, 2011. Department of Computer Science Center for Visual Computing ICPR 2012 Background (1) Heat Kernel on surface 𝑀 Amount of heat transferred from a point 𝑥 to 𝑦 in time 𝑡 −𝜆𝑙 𝑡 ℎ𝑡 𝑥, 𝑦 = ∞ 𝜙𝑙 𝑥 𝜙𝑙 (𝑦) 𝑙=0 𝑒 𝜆𝑙 , 𝜙𝑙 : 𝑙-th eigenvalue and eigenfunction of the Laplace-Beltrami operator Heat Kernel Signature (HKS): ℎ𝑡 (𝑥, 𝑥) HKS descriptor A vector of HKS probed at different values of 𝑡 𝐾 𝑥 = (ℎ𝑡 1 𝑥, 𝑥 , ℎ𝑡2 𝑥, 𝑥 , … , ℎ𝑡𝑟 𝑥, 𝑥 ) Properties of Heat Kernel Intrinsic (Invariant to rigid and isometric deformation) Informative (locally and globally shape aware) Stable Department of Computer Science Center for Visual Computing ICPR 2012 Background (2) Geometric words 𝑊𝑖 A representative HKS vector Clustered in the HKS descriptor space by the k-means algorithm Vocabulary 𝑾 = {𝑊1 , … , 𝑊𝑉 } Similarity of point 𝑥 and word 𝑊𝑖 𝜃𝑖 𝑥 = 𝑐 𝑥 Department of Computer Science Center for Visual Computing 𝐾 𝑥 −𝑊𝑖 2 − 2𝜎2 𝑒 ICPR 2012 Shape-Google Revisit (1) Bag-of-Words Word distribution of each point Θ 𝑥 = 𝜃1 𝑥 , … , 𝜃𝑉 𝑥 BoW descriptor: 𝑉 × 1 vector 𝑓 𝑀 = Θ 𝑥 𝑑𝜇(𝑥) 𝑀 Measure the frequencies of words appearing on the shape Department of Computer Science Center for Visual Computing ICPR 2012 Shape-Google Revisit (2) Spatially-Sensitive Bag-of-Words SS-BOW descriptor: 𝑉 × 𝑉 matrix 𝐹 𝑀 = 𝑀×𝑀 Θ 𝑥 ΘT y ℎ𝑡 𝑥, 𝑦 𝑑𝜇 𝑥 𝑑𝜇(𝑦) Measure the frequencies of word pairs Department of Computer Science Center for Visual Computing ICPR 2012 New Paradigm: Bag-of-Feature Graphs (1) Motivation: Reduce computation complexity Considering all points on shape -> only considering feature points Vector/matrix of word frequencies -> feature graphs associated with words Department of Computer Science Center for Visual Computing ICPR 2012 Formulation (1) Feature set: 𝐹 Feature graph 𝐺𝑖 associated with the 𝑖-th geometric word 𝑊𝑖 𝐺𝑖 represented as 𝐹 × |𝐹| matrix 𝐺𝑖 𝑥, 𝑦 = 𝜃𝑖 𝑥 𝜃𝑖 𝑦 ℎ𝑡 (𝑥, 𝑦) 𝑥, 𝑦 ∈ 𝐹 ℎ𝑡 (𝑥, 𝑦): Heat Kernel Bag-of-Feature-Graphs representation of shape 𝑀 𝐺 𝑀 = {𝐺1 , 𝐺2 , … , 𝐺𝑉 } … 𝐺1 Department of Computer Science Center for Visual Computing 𝐺2 ICPR 2012 𝐺3 Formulation (2) BoFG descriptor Multi-dimensional scaling (MDS): Choosing the 6 largest eigenvalues of each graph matrix denoted by 𝑆𝑖 𝑀 6𝑉 × 1 vector 𝑆1 𝑀 , … , 𝑆𝑉 𝑀 𝑇 Shape distance 𝑉 𝑑 𝑀1 , 𝑀2 = 𝑆𝑖 𝑀1 − 𝑆𝑖 𝑀2 2 𝑖=1 Retrieval by approximate nearest neighbor (ANN) search Department of Computer Science Center for Visual Computing ICPR 2012 Nonrigid Shapes and Their BoFG Descriptors Department of Computer Science Center for Visual Computing ICPR 2012 𝑁: Number of vertices 𝑂 𝐷 : Time complexity for computing HKS descriptor of a vertex Time Complexity of BoW, SS-BOW and BoFG Department of Computer Science Center for Visual Computing ICPR 2012 Experiments Test dataset: TOSCA1 12 classes of 148 non-rigid shapes Each shape has 3K ∼ 30K vertices Evaluated methods: BoW, FSS-BoW, SI-HKS, Vocabulary 48 words for BoW and SS-BoW (clustered from all shape points) 4 words for BoFG (clustered only from feature points) Feature numbers in BoFG: 30 ∼ 50 for each shape 1http://toca.cs.technion.ac.il/book/shrec.html Department of Computer Science Center for Visual Computing ICPR 2012 Department of Computer Science Center for Visual Computing ICPR 2012 Experiments Time performance (in seconds) of three descriptors on two shapes with 3K and 30k vertices Department of Computer Science Center for Visual Computing ICPR 2012 Experiments (1) (2) (3) Precision-recall curves of evaluated methods, with categories of (1) null, (2) scale changes and (3) holes. Department of Computer Science Center for Visual Computing ICPR 2012 Partial shape retrieval Query shape is only a part of a complete model Online feature alignment is required to extract corresponding subgraphs Department of Computer Science Center for Visual Computing ICPR 2012 Summary Bag-of-Feature-Graphs (BoFG) is a new paradigm for shape representation This representation is feature-driven, concise, and spatially-aware The key idea is to construct graphs of features associated with geometric words BoFG has much improved time-performance and competitive retrieval results in comparison with other state-of-the-art methods Department of Computer Science Center for Visual Computing ICPR 2012 Future Work Investigate graph comparison with heavy outliers Improve the performance on partial shape retrieval Acknowledgements: Research Grants from National Science Foundation Department of Computer Science Center for Visual Computing ICPR 2012 Thank You! Department of Computer Science Center for Visual Computing ICPR 2012