KITTI-trajectory-prediction

KITTI is a well established dataset in the computer vision community. It has often been used for trajectory prediction despite not having a well defined split, generating non comparable baselines in different works. This dataset aims at bridging this gap and proposes a well defined split of the KITTI data. Samples are collected as 6 seconds chunks (2seconds for past and 4 for future) in a sliding window fashion from all trajectories in the dataset, including the egovehicle. There are a total of 8613 top-view trajectories for training and 2907 for testing. Since top-view maps are not provided by KITTI, semantic labels of static categories obtained with DeepLab-v3+ from all frames are projected in a common top-view map using the Velodyne 3D point cloud and IMU. The resulting maps have a spatial resolution of 0.5 meters and are provided along with the trajectories.