IDD-3D

自动驾驶和辅助系统依靠来自交通和道路场景的注释数据来建模和学习复杂现实场景中的各种对象关系。可部署的深度学习体系结构的准备和培训要求模型适合不同的流量场景并适应不同的情况。当前，现有的数据集虽然规模很大，但缺乏这种多样性，并且在地理上偏向于主要发达的城市。在印度等几个发展中国家发现的非结构化和复杂的驾驶布局对这些模型提出了挑战，因为对象类型，密度和位置的变化程度很大。为了促进更好的研究以适应此类场景，我们构建了一个新的数据集，IDD-3D，该数据集由来自多个摄像机的多模式数据和LiDAR传感器组成，该传感器具有跨各种交通场景的12k带注释的驱动LiDAR帧。我们通过与现有数据集的统计比较来讨论对该数据集的需求，并突出显示复杂布局中标准3D对象检测和跟踪任务的基准。Autonomous driving and advanced driver-assistance systems rely on annotated data from traffic and road scenes to model and learn the relational patterns of various objects in complex real-world scenarios. The development and training of deployable deep learning architectures require models to adapt to diverse traffic scenarios and varying conditions. Currently, although existing datasets are large in scale, they lack such diversity and are geographically biased towards major developed urban areas. Unstructured and complex driving layouts found in several developing countries such as India pose challenges for these models, as the variability in object types, densities, and positions is substantial. To facilitate better research on adapting to such scenarios, we construct a novel dataset, IDD-3D, which consists of multimodal data from multiple cameras and LiDAR sensors, containing 12,000 annotated driving LiDAR frames spanning diverse traffic scenarios. We discuss the necessity of this dataset through statistical comparisons with existing datasets, and highlight benchmarks for standard 3D object detection and tracking tasks in complex driving layouts.

搜集汇总

背景与挑战

背景概述

IDD-3D是一个用于自动驾驶研究的数据集，针对非结构化驾驶场景（如印度）的多样性不足问题而构建。它包含多模态数据和LiDAR传感器，提供了12k带注释的LiDAR帧，以支持复杂交通场景下的3D对象检测和跟踪任务。