Social Human Pose Dynamics and Trajectory Forecasting results derived from the PoseTrack dataset measured using at several future frames
|Method||Score||80 ms||160 ms||320 ms||400 ms||560 ms|
Social Human Pose Dynamics and Trajectory Forecasting results derived from the 3DPW dataset measured using VIM at several future frames
|Method||Score||100 ms||240 ms||500 ms||640 ms||900 ms|
The scores reported in the first column of these leaderboards are calculated by first averaging the scores from the other five columns and then converting these values to z-scores based on the resulting distribution. These scores are then normalized so the best-performing method receives a score of 1 and the lowest-performing method receives a score of 0. These scores may change if new methods are submitted.
|VIM (Visibility-Ignored Metric)||Average Euclidean distance between ground-truth and estimated joint positions when invisible joints are discarded|
|VAM (Visibility-Aware Metric)||β if visibility is predicted incorrectly and the minimum of β and the distance between the ground-truth and esimated positions otherwise|
Note that the scores reported on the leaderboard are scaled from meters to centimeters for 3DPW. We use a β value of 200 for calculating VAM.