Dataset Introduction:
Welcome to the MSA-BUPT, in order to compensate for the shortcomings of the existing datasets, a new high-quality multi-situation video person ReID dataset, named MSA-BUPT, is built to promote the video person ReID in large-scale urban surveillance. Specifically, compared to other existing datasets, MSA-BUPT contains a variety of scene variation, viewpoint variation, and person status variation.
Specifically, MSA-BUPT contains the following variations:
(1) Lighting variation: The pedestrian videos are captured in both morning and afternoon, so there will be a noticeable change in lighting in the images.
(2) Scene variation: To ensure that pedestrian is captured in the most realistic way as possible, we design a variety of scenes such as indoor corridors, lobbies, staircases, outdoor squares, green belts, crossroads, streets, and corners for the MSA-BUPT dataset. As a consequence, multiple forms of interactive information from different environments could be obtained.
(3) Viewpoint variation: We design a surveillance network with cameras of various height, i.e., 1.8m mobile cameras, 3m indoor surveillance cameras and 5m outdoor surveillance cameras. Since the cameras cover a wide area, the foreshortening effects for pedestrians are very clear, making the dataset suitable for real-world applications.
(4) Recording angle variation: To ensure the representation of various viewing angles, the camera network is designed to take the curves and turns of most roads into consideration. Therefore, the capture of a continuous trajectory including front, side, and back images of the pedestrian, as well as multiple postures of the pedestrian could be performed smoothly, leading to a multi-angle view of the subject.
(5) Scale variation: In MSA-BUPT, a part of the trajectory contains the entire process of the pedestrian walking towards or away from the camera. There is a wide range of scales in the dataset, from 70- 600px in height and 20-300px in width, reflecting differences in scale observed by the camera and applicable to indoor environments as well. Moreover, the images within this dataset have various aspect ratios, which are beneficial for testing the robustness of the video person ReID algorithm.
(6) Clothing and posture variations: The data of MSA-BUPT is captured at the turn of seasons, resulting in a diversity of clothing companying with a variety of postures, such as walking, running, interacting, and carrying heavy objects, etc. Moreover, the pedestrian wearing masks occurs in over 50% of the dataset, which benefits for coarse screening in pedestrian tracking. Attribute annotations: To facilitate further research, we annotate person attribute, including appearances, clothing, and accessories. It is beneficial to other tasks such as multimodal analysis.
地址:北京市西土城路10号 邮编:100876 京ICP备 18055448号-1
