A Realistic Predictor for Pedestrian Attribute Recognition

The application of video surveillance systems in public areas to ensure public security is becoming increasingly important. A major task when evaluating the arising amount of video data is to find the occurrences of a person-of-interest on the basis of a testimony. For the comparison of a person's description with persons in the video data, the attributes of all persons must be recognized automatically. However, typical approaches to pedestrian attribute recognition simply predict all attributes for a person, regardless the visibility of relevant attributes. To address this problem, the concept of realistic predictors is used in this work to determine and improve the reliability of pedestrian attribute recognition.