Salient feature based graph matching for person re-identificationPattern Recognition


Sara Iodice, Alfredo Petrosino
Artificial Intelligence / Computer Vision and Pattern Recognition / Signal Processing / Software


Salient Color Names for Person Re-identification

Yang Yang, Jimei Yang, Junjie Yan, Shengcai Liao, Dong Yi, Stan Z. Li

Transfer re-identification: From person to set-based verification

Wei-Shi Zheng, Shaogang Gong, Tao Xiang

Multi-shot SURF-based person re-identification via sparse representation

Mohamed Ibn Khedher, Mounim A. El Yacoubi, Bernadette Dorizzi

Improved Person Re-Identification Using Statistical Approximation

Yan Yang, Farhad Dadgostar, Sandra Mau, Brian C. Lovell

Multi-view implicit transfer for person re-identification

Wei Xu, Yijun Li, Chen Gong, Jie Yang


Salient feature based graph matching for person re-identification

Sara Iodice 1, Alfredo Petrosino n,1

Department of Science and Technology, University of Naples Parthenope, Centro Direzionale Is C4, 80143 Naples, Italy a r t i c l e i n f o

Article history:

Received 2 January 2014

Received in revised form 12 September 2014

Accepted 15 September 2014


Computational symmetry

Salient features

Graph representation and matching

People re-identification a b s t r a c t

We propose a person re-identification non-learning based approach that uses symmetry principles, as well as structural relations among salient features. The idea comes from the consideration that local symmetries, at different scales, also enforced by texture features, are potentially more invariant to large appearance changes than lower-level features such as SIFT, ASIFT. Finally, we formulate the re-identification problem as a graph matching problem, where each person is represented by a graph aimed not only at rejecting erroneous matches but also at selecting additional useful ones.

Experimental results on public dataset i-LIDS provide good performance compared to state-of-theart results. & 2014 Elsevier Ltd. All rights reserved. 1. Introduction

Symmetry detection is highly relevant in pattern recognition.

Indeed, the description of a figure may be different when it is embedded in a context with horizontal or vertical symmetry [19].

Besides, in tasks requiring the completion of partially occluded visual stimuli, subjects tend to produce systematically symmetrical [14]. The concept of symmetry is not univocal: various kinds of properties of an image are defined as symmetry [30,28].

As instance, a figure has rotational symmetry when it can be rotated less than 3601 around its central point, or axis, and still matches the original figure.

This cue is peculiar in person re-identification where the problem consists in recognizing people in different poses from images coming from distinct cameras. This is an important task in the video surveillance, where large and structured environments must be supervised (such as airport, metro, station or shopping centres) and it becomes more critical when the cardinality of gallery set increases.

Symmetry is adopted in [6], a person re-identification method that weighs appearance information, extracted from different body parts, in accordance with their distance from symmetry axes computed on the whole figure. Therefore, symmetry appears to have been used as a global property on this previous work, not as a local one. Conversely, symmetry is adopted as both global and local property in our approach. In particular, the global symmetry axis is exploited to select salient locations representing each pedestrian, while local symmetry is adopted like local feature in order to describe each salient location.

Researchers in computer vision have made significant progress in representing and detecting symmetries in images and other types of data [15].

However, there has been relatively little work on using local symmetries as explicit features for matching tasks [2], and no work about matching the same pedestrian with different poses, like in re-identification. The main idea is to use a variety of symmetries, rather than repetitions, together with texture as cues, in order to define a complex feature detector, based on the consideration that, at different scales, the symmetry is potentially more invariant to large appearance changes than lower-level features such as SIFT, and, when combined with a texton-based feature, is highly discriminative [11].

In this context, the main contribution resides in the image features and the adopted feature organization for matching. The discriminative power could be more appreciated if the features are organized in an appropriate manner. Most of the researchers focus their attention on the features extracted in many ways from the scene. The Harris and SIFT [16] features are important to identify what is distinctive and discriminative for the purpose of a correct recognition of the scene. The bag of words algorithm has been applied to SIFT descriptors, to identify discriminative combinations of descriptors [4]. In [18] the application of a clustering to descriptors leads to results which are less distinctive in a large cluster than those in a small cluster. For example in indoor navigation, window corners are common, so they are not good features to uniquely identify scenes, while corners found on posters or signs are much better. In [7] an effective approach based on real-time loop detection has proved to be efficient using a hand-held camera, through SIFT features and intensity and hue

Contents lists available at ScienceDirect journal homepage:

Pattern Recognition 0031-3203/& 2014 Elsevier Ltd. All rights reserved. n Corresponding author.

E-mail addresses: (S. Iodice), (A. Petrosino). 1 Tel.:/fax: þ39 0815476656.

Please cite this article as: S. Iodice, A. Petrosino, Salient feature based graph matching for person re-identification, Pattern Recognition (2014),

Pattern Recognition ∎ (∎∎∎∎) ∎∎∎–∎∎∎ histograms combined using a bag of words approach. Up to now none of the existing approaches tackled the relations among features in terms of similarity and spatial homogeneity.

Our main contribution consists in introducing an approach based on graph-based representation, according to which regions with their corresponding feature vector and the geometric relationship between these regions are encoded in the form of a graph. According to the idea that an image can be described at the higher level in terms of a nested hierarchy of local symmetries, we present a novel graph matching approach to the problem aimed at evolving an initial set of correspondences computed with the local features, as a kind of compromise between the constraints imposed by both the local features and the structural relations.