Adaptive Multi-Metric Fusion for Person Re-identification
Person re-identification, which aims at recognizing a person of interest across spatially disjoint camera views, is still a challenging task. Plenty of approaches emerge in recent years and some of them achieve good matching results. Given a probe image,
- PDF / 1,688,848 Bytes
- 10 Pages / 439.37 x 666.142 pts Page_size
- 22 Downloads / 205 Views
Institute of Image Processing and Pattern Recognition, Shanghai Jiao Tong University, 800, Dongchuan Road, Shanghai 200240, China {lipenglin,liumengxue,lxyao,jieyang}@sjtu.edu.cn 2 School of Biomedical Engineering, Shanghai Jiao Tong University, 800, Dongchuan Road, Shanghai 200240, China [email protected] Abstract. Person re-identification, which aims at recognizing a person of interest across spatially disjoint camera views, is still a challenging task. Plenty of approaches emerge in recent years and some of them achieve good matching results. Given a probe image, we observe that the ranking results generated by different approaches differ from each other. Considering these conventional methods are reasonable, we propose an Adaptive Multi-Metric Fusion (AMMF) method which fuses the existing ranking results with query-specific weights. Experiments on two challenging databases, VIPeR and ETHZ, demonstrate that the proposed method achieves further performance improvement.
Keywords: Person re-identification Re-ranking
1
·
Fusion
·
Multi-metric
·
Introduction
The fundamental task of person re-identification is to recognise and associate a person over a distributed surveillance network. In general, we simplify this task as matching each probe image with a gallery data set composed of numerous candidates, which can be viewed as a special case of image retrieval. This problem is challenging due to camera variations, low image resolutions, changing poses, viewpoints and illumination conditions [1–4]. To tackle such challenges, current research efforts mainly focus on two aspects: feature representation methods and metric learning methods. Feature representation methods concentrate on designing discriminative descriptors which are robust to changes in viewpoint, illumination and background clutters. Some effective person image descriptors include kBiCov [5], salience match [6], SCNCD [7], query-adaptive fused feature [8] and LOMO [9]. Metric learning methods aim to learn a distance metric which measures the similarity between person images. Under the learned optimal metric, the intra-class distances are minimized while the inter-class distances are maximized. Representative learning methods include Information Theoretic Metric Learning (ITML) [10], Logistic c Springer Nature Singapore Pte Ltd. 2016 T. Tan et al. (Eds.): CCPR 2016, Part I, CCIS 662, pp. 258–267, 2016. DOI: 10.1007/978-981-10-3002-4 22
Adaptive Multi-Metric Fusion for Person Re-identification
259
Fig. 1. Given a probe image xip , its correct matching gallery image xjg is ranked at different positions (marked with green border) by 3 different re-identification methods. Obviously, Method #1 is more discriminative for gallery image xjg . (Color figure online)
Discriminant Metric Learning (LDML) [11], Large Margin Nearest Neighbors (LMNN) [12] and Keep It Simple and Straightforward (KISSME) [13], etc. In this paper, we investigate the ranking results obtained from previous methods, finding that different methods get different ranking results for a specific pro
Data Loading...
