The continuously growing amount of available image and video data of public spaces provides new opportunities in public safety and law enforcement. Manual sifting is difficult due to the large amount of data.
Automated person re-identification can greatly accelerate and alleviate the evaluation and navigation in the data material. Typically, re-identification methods attempt to decide if two person images show the same or different persons. This enables person search in big data scenarios.
Persons and their semantic attribute descriptions. Image source: Wang, X. et al.: Pedestrian Attribute Recognition: A Survey, 2019
The drawback of these methods is that they are only operational if image material of the target person is present. In practice, however, often only witness descriptions of a person are available. For this reason, the attribute-based person re-identification deals with person search solely based on semantic descriptions of a person’s attributes, e. g. the clothing or gender.
Deep learning approaches are developed which can identify persons in a large amount of surveillance data merely by means of a semantic description of a person. The difficulty of different modalities of textual query and image data is addressed as well as different characteristics of attributes and occlusions.
Semantic attribute query and resulting rank list of surveillance images. Image source: Zheng, L. et al.: “Scalable Person Re-Identification: A Benchmark“. In Proceedings of the IEEE International Conference on Computer Vision (ICCV), 2015