People search with textual queries on clothing appearance

People search with textual queries on clothing appearance is a recently introduced computer vision task, related to person re-identification. It consists of retrieving images of individuals that match a textual query describing the clothing appearance (e.g., “individuals wearing a white shirt and blue trousers”). This can be useful, e.g., in forensic video analysis, when only textual descriptions of individuals of interest, given by witnesses, can be available.

We developed a people search method based on our dissimilarity descriptor of clothing appearance, originally developed for re-identification tasks, with the aim of extending any given re-identification system. Basically, we identify basic high-level clothing appearance characteristics ("attributes") that can be detected using the low-level descriptor at hand (e.g., the color and texture of upper and lower body garments); we then implement a detector for each attribute as a binary classifier that outputs a score related to the likelihood that the corresponding attribute is present in the input image (e.g., a support vector machine). Queries made up of a boolean combination of several attributes, like "white shirt AND blue trousers", can be conveniently processed using fuzzy logic operators.