This dataset contains two experiments coined as ETH (left image) and Hotel (right image):
- This dataset is extensively used in Human Trajectory Prediction literature.
Each directory contains a video together with the annotation and the obstacle map used. Please read the following for some details about the provided files. In each sequence directory there is a info.txt file with some sequence specific information.
The actual annotation is stored in the obsmat.txt file. Each line has this format
[frame_number pedestrian_ID pos_x pos_z pos_y v_x v_z v_y ]
however pos_z and v_z (direction perpendicular to the ground) are not used. The positions and velocities are in meters and are obtained with the homography matrix stored in H.txt .
Please note that we tried to avoid annotating those subjects that were at the border of the scene, as their behavior might have been influenced by the presence of other pedestrians/obstacles not in the field of view of the camera. We tried to be as consistent as possible in this regard.
We tried to keep note of the people that seemed to walk in groups. These are listed in the file groups.txt . Each line contains a list of id, that are those that form a group. The id are the same as those in the obsmat.txt file
The assumed destinations for all the subjects walking in the scene are stored in the file destinations.txt . This is of course a simplifyiing assumption, but it seems to work fine for us.
the obstacles are reported in the map.png file. To bring the obstacle from image to world coordinates, the use of the homography matrix is necessary.
- WARNING: on 17/09/2009 the dataset have been modified, the frame number in the obsmat had a wrong offset (Thanks for corrections to Paul Scovanner)
Homography matrics are provided by dataset creators in H.txt files. You can use them to project the world-coord positions (or trajectories) on the reference images (or videos).
Below is a sample python function to transform a trajectory from world coordinate system to image pixels:
import os
import numpy as np
def world2image(traj_w, H_inv):
# Converts points from Euclidean to homogeneous space, by (x, y) → (x, y, 1)
traj_homog = np.hstack((traj_w, np.ones((traj_w.shape[0], 1)))).T
# to camera frame
traj_cam = np.matmul(H_inv, traj_homog)
# to pixel coords
traj_uvz = np.transpose(traj_cam/traj_cam[2])
return traj_uvz[:, :2].astype(int)
H = (np.loadtxt(os.path.join(OPENTRAJ_ROOT, "datasets/ETH/seq_eth/H.txt")))
H_inv = np.linalg.inv(H)
world2image({TRAJ}, H_inv) # TRAJ: Tx2 numpy arrayIn order to the load the datasets, we provided the loader_eth.py
import os
from toolkit.loaders.loader_eth import load_eth
# fixme: replace OPENTRAJ_ROOT with the address to root folder of OpenTraj
traj_dataset = load_eth(os.path.join(OPENTRAJ_ROOT, "datasets/ETH/seq_eth/obsmat.txt"))
trajs = traj_dataset.get_trajectories()No license information is available with this dataset.
@inproceedings{pellegrini2009you,
title={You'll never walk alone: Modeling social behavior for multi-target tracking},
author={Pellegrini, Stefano and Ess, Andreas and Schindler, Konrad and Van Gool, Luc},
booktitle={2009 IEEE 12th International Conference on Computer Vision},
pages={261--268},
year={2009},
organization={IEEE}
}
- Please write to [email protected] for any question or comment.