KITTI-trajectory-prediction

KITTI 是计算机视觉社区中成熟的数据集。尽管没有明确定义的分割，但它经常被用于轨迹预测，在不同的工作中生成不可比较的基线。该数据集旨在弥合这一差距，并提出了对 KITTI 数据的明确划分。以滑动窗口的方式从数据集中的所有轨迹（包括 egovehicle）中以 6 秒的块（过去 2 秒，未来 4 秒）的形式收集样本。共有 8613 条顶视图轨迹用于训练，2907 条用于测试。由于 KITTI 不提供顶视图，使用 Velodyne 3D 点云和 IMU 将使用 DeepLab-v3+ 从所有帧中获得的静态类别的语义标签投影到一个通用的顶视图中。生成的地图具有 0.5 米的空间分辨率，并与轨迹一起提供。

KITTI is a well-established dataset in the computer vision community. Although no clearly defined splits exist for it, KITTI is frequently utilized for trajectory prediction, resulting in incomparable baselines across different research works. This dataset aims to bridge this gap and proposes a clear split for the KITTI data. Samples are collected in the form of 6-second chunks (2 seconds past and 4 seconds future) from all trajectories (including the egovehicle) in the dataset via a sliding window approach. A total of 8613 bird's-eye view trajectories are used for training, while 2907 are reserved for testing. Since KITTI does not provide bird's-eye view data, Velodyne 3D point clouds and IMU are employed to project semantic labels of static categories extracted from all frames via DeepLab-v3+ onto a unified bird's-eye view. The generated maps have a spatial resolution of 0.5 meters and are provided together with the trajectories.

搜集汇总

背景与挑战

背景概述

该数据集基于KITTI，专门为轨迹预测任务提供明确的训练和测试划分，包含8613条训练轨迹和2907条测试轨迹。数据以6秒滑动窗口（过去2秒，未来4秒）形式组织，并利用Velodyne点云和IMU将语义标签投影到顶视图，生成0.5米分辨率的地图与轨迹。