[1H5-GS-10] Vision, speech: Recognition and detection

[1H5-GS-10-04] Speeding up inference by simultaneous execution of person detection and feature extraction in Person Search

〇Kento Hori¹, Masahiro Suzuki¹, Yutaka Matsuo¹ (1. University of Tokyo)

Keywords:Computer Vision, Person Search, Person Re-Identification

Person Search is a task that predicts whether a person is a particular person from images obtained from cameras, and has been actively researched because of its high applicability in the real world, such as tracking and searching for people.
In the past research, the focus was on the accuracy of the proposed method, whether it is used in the field such as business, etc. In this paper, we discuss the method, and it is necessary to emphasize efficiency as an index for evaluating the Person Search method.
In terms of efficiency, in the conventional method, the feature vector of the person is acquired after detecting the person, and the processing is redundant, so in this study, the feature extraction of the image was performed to improve the efficiency. We propose a method to output the human region and the human feature vector simultaneously and execute two tasks in parallel by devising the structure of the prediction part.
Experiment on the PRW (Person Re-identification in the Wild) dataset was performed to confirm that the proposed method achieved the same level of accuracy with a smaller number of parameters and a faster inference speed than the conventional method.