Person Re-identification using Co-occurrence Attributes of Physical and Adhered Human Characteristics

We combine physical characteristics, such as gender and age, with associated appearance characteristics, such as clothing, to form co-occurrence attributes. Classifier confidence scores for these attributes serve as features for matching the same person across different images.

Person re-identification means finding whether people shown in different images are the same individual. We proposed using combinations of characteristics rather than treating gender, age, hairstyle, and clothing only as separate features. Examples include a man wearing a suit or a person with long hair wearing a skirt.

We call these combinations co-occurrence attributes and designed them with reference to methods used by public websites for searching for people. First, we train classifiers to judge each combination from an image. Their confidence scores form a list of numbers describing the person. A method that learns image similarity then compares this list with those from reference images.

Experiments on several public datasets showed a substantial improvement in matching performance over the existing methods used for comparison. We also demonstrated how to analyze which combinations were most useful. Combining characteristics provides a richer description for distinguishing people across images.

• Masashi Nishiyama, Shota Nakano, Tappei Yotsumoto, Hiroki Yoshimura, Yoshio Iwai and Kazunori Sugahara,
  Person Re-identification using Co-occurrence Attributes of Physical and Adhered Human Characteristics, Proceedings of 23rd International Conference on Pattern Recognition (ICPR), pp. 2086 - 2091, December 2016.

• 中野 翔太, 四元 辰平, 吉村 宏紀, 西山 正志, 岩井 儀雄, 菅原 一孔,
  身体と外見の共起属性を用いた人物対応付け,
  電子情報通信学会論文誌 D, Vol.J100-D, No.1, pp.104-114, January 2017.