KinGaitWild (Kinship verification from gait dataset)

To study kinship verification from gait, we collected the dataset KinGaitWild consisting of several videos from youtube. Most of the videos were taken under uncontrolled conditions in terms of background, camera motion, luminance and viewpoints. The KinGaitWild dataset contains 105 videos of celebrities and their relatives. The average time duration of each video is around 10 seconds. The database includes 60 pairs of Father-Son (FS) relationships. These pairs are equally split into 5 groups. We focus in this study on the Father-Son relationships. The database collection was done as follows. First, we used the YouTube Data API to search for videos showing celebrities walking in the wild. To avoid biases, we selected the pairs of celebrities so that the videos are not originated from the same source nor environment. For each video, we labeled the position of each specified person by a bounding box (bbox). These bboxes are used to estimate the human pose for silhouette-based approaches and to crop the region of interest for both the video-based and deep-learning-based approaches.