Ithaca365

近年来，由于通常在特定位置和良好天气条件下收集的大规模数据集的可用性，自动驾驶汽车的感知进步加速了。然而，为了实现高安全性要求，这些感知系统必须在包括雪和雨在内的各种天气条件下稳健运行。在本文中，我们提出了一个Ithaca365，一个新的数据集，通过一种新颖的数据收集过程实现稳健的自动驾驶-数据是在不同的场景 (城市，高速公路，农村，校园)，天气 (雪，雨，太阳) 下沿着15千米路线重复记录，时间 (白天/晚上) 和交通状况 (行人、骑自行车的人和汽车)。该数据集包括来自四个摄像机和激光雷达传感器的图像和点云，以及高精度GPS/INS，以建立跨路线的对应关系。数据集包括使用amodal遮罩捕获部分遮挡和2D/3D边界框的道路和对象注释。我们通过分析基线在道路和对象的平均分割，深度估计和3D对象检测中的性能来证明该数据集的独特性。重复的路线为对象发现，持续学习和异常检测开辟了新的研究方向。In recent years, advances in autonomous vehicle perception have accelerated due to the availability of large-scale datasets typically collected at specific locations under favorable weather conditions. However, to meet stringent safety requirements, these perception systems must operate robustly across diverse weather conditions including snow and rain. In this paper, we present Ithaca365, a novel dataset for robust autonomous driving enabled by an innovative data collection pipeline. Data was repeatedly recorded along a 15-kilometer route across various scenarios (urban, highway, rural, campus), weather conditions (snow, rain, sunshine), times of day (daytime/nighttime), and traffic states (pedestrians, cyclists, and motor vehicles). This dataset includes images and point clouds from four cameras and LiDAR sensors, as well as high-precision GPS/INS to establish cross-route correspondences. It also features road and object annotations: amodal masks are used to capture partial occlusions, alongside 2D/3D bounding boxes for both road structures and objects. We demonstrate the uniqueness of this dataset by evaluating baseline performance on road and object semantic segmentation, depth estimation, and 3D object detection tasks. The repeated route collection also opens up new research avenues for object discovery, continual learning, and anomaly detection.

搜集汇总

背景与挑战

背景概述

Ithaca365是一个专注于自动驾驶感知的数据集，通过在不同场景、天气和时段下重复收集15千米路线的数据，包含图像、点云和注释，旨在提升系统在恶劣条件下的鲁棒性。它为对象发现、持续学习等研究提供了新方向。