mining discriminative co-occurrence patterns for visual recognition - yuan, yang, wu - proceedings...

Mining Discriminative Co-occurrence Patterns for Visual Recognition Junsong Yuan School of EEE Nanyang Technological University Singapore 639798 [email protected] Ming Yang Dept. of Media Analyti cs NEC Laboratories America Cupertino, CA, 95014 USA [email protected] Ying Wu EECS Dept. Northwestern University Evanston, IL, 60208 USA [email protected] Abstract The co-occurrence pattern, a combination of binary or local features, is more discriminative than individual features and has shown its advantages in object, scene, and action recognition. We discuss two types of co-occurrence patter ns that are compleme ntary to each other , the con- junction (AND) and disjunction (OR) of binary features. The necessary condition of identifying discriminative co- occurren ce patterns is firstly provid ed. Then we pr opose a novel data mining method to efficiently discover the op- timal co-occurrence pattern with minimum empirical error, despit e the noisy train ing dataset. This mining proc edur e of AND and OR patterns is readily integrated to boosting, which improves the generalization ability over the conven- tion al boosti ng dec isi on tr ees and boos ting dec isi on stu mps . Our versatile experiments on object, scene, and action cat- egorization validate the advantages of the discovered discriminative co-occurrence patterns. 1. Introduction Due to the compos itional proper ty of vis ual obj ect s, scene s, and actions, the discover y of disc rimi nati ve co- occurr ence pattern is of fundamental impor tance in rec- ognizi ng them. Altho ugh the extrac ted featu res, such as col or , te xture, sha pe, or mot ion features, can be qui te weak individually, an appropriate combination of them will bri ng a str ong fea tur e whi ch is muc h mor e dis cri min a- tive [31] [29] [3] [16] [2] [9]. Ther e has been a recent trend in mining co-occurrence patterns for visual recognition. For exampl e, eve ry real-world object is associ ated with numerous visua l attr ibut es in terms of its materia l, structure, shape, etc, [11] [4][10], although it is difficult to differentiate them using a single visual attrib ute, they can be well distinguished by the co-occurrence of specific attributes, as illustrated in Fig. 1. In a binar y class ifica tion problem, giv en a colle ctio n of N binar y feat ures, the problem of co-oc curre nce pat- te rn mi ni ng is to se le ct a subset fr om thes e N fea- tur es, suc h tha t the co-occ urr enc e of the m can best dis - Figure 1. By inferring binary visual attributes from raw image features, such as wheels, furry [6] [4], we can distingui sh bikes from people by a co-occurrence of certain attributes, for example a bike has metal and wheels, but does not have head . Given two classes of objects, described by (possibly quite noisy) attributes, can we efficiently discover the co-occurrence of attributes that can best discriminate them ? cr imina te the two cl a ss es . In sp it e of ma ny p revi - ous works in mini ng and inte grati ng co-occurre nce patterns [22] [20] [3] [29] [33][16] [24] [9] [27], none of these methods is targeted at finding the most discriminative co-occurrence pattern with the smallest classification error . Given N binary features, because the co-occurrence pattern can contain an arbitrary number of features (up to N ), the total number of candidates of co-occurrence patterns is exponentially large ( e.g. 3 N if considering the neg- ative value or 2 N if only considering the positive value.) As a resul t, it is compu tati onall y intracta ble to perfo rm an exhaus tiv e search . Even wors e, unlik e conve ntiona l feat ure sele ction task, the monot onic property of feat ure subset does not hold in searching co-occurrence patterns. Namely, a (K + 1) -order binary feature is not necessarily better tha n a K -orde r one. There fore, the branch -and-b ound search cannot be applied directly [18]. Existing approaches for co-occurre nce pattern search, such as seque ntial for- ward selection [16] [24], or recent data mining-driven approaches [22] [20] [3] [29] [33], do not guarantee the op- timality of the selected co-occurre nce patterns. In general, when the training data is noisy, it is still an open problem to find the co-occurrence pattern of the bes t performance, e.g., minimum classification error [17]. 2777

mining discriminative co-occurrence patterns for visual recognition - yuan, yang, wu - proceedings...

Documents