人工智能培训

搜索

人工智能论文:APOLLOCAR3D:一个大型3D汽车实例,了解自动驾驶的基准(ApolloCar3D: A Large 3D Car Instance Und

[复制链接]
nic28 发表于 2018-11-30 10:24:24 | 显示全部楼层 |阅读模式
nic28 2018-11-30 10:24:24 3228 0 显示全部楼层
人工智能论文:APOLLOCAR3D:一个大型3D汽车实例,了解自动驾驶的基准(ApolloCar3D: A Large 3D Car Instance Understanding Benchmark for  Autonomous Driving)自动驾驶引起了业界和学术界的极大关注。一个重要的任务是估计道路上移动或停放的车辆的3D特性(例如,翻译,旋转和形状)。这项任务虽然具有批判性,但在计算机视觉领域仍未得到充分研究 - 部分原因在于缺乏适用于自动驾驶研究的大规模和完全注释的3D汽车数据库。在本文中,我们提供了适用于3D汽车实例理解的第一个大型数据库-ApolloCar3D。该数据集包含5,277个驾驶图像和超过60,000个车辆,其中每辆车都配备了具有绝对模型尺寸和语义标记关键点的行业级3D CAD模型。该数据集比PASCAL3D +和KITTI(现有技术水平)大20倍以上。为了在3D中实现高效的标记,我们通过考虑单个实例的2D-3D关键点对应关系和多个实例之间的3D关系来构建管道。配备这样的数据集,我们使用最先进的深度卷积神经网络构建各种基线算法。具体来说,我们首先使用预先训练的Mask R-CNN对每辆车进行分段,然后根据其3D姿态和形状进行恢复。一个可变形的3D汽车模型,没有使用语义关键点。我们表明,使用关键点可以显着提高拟合性能。最后,我们开发了一个新的3D度量,共同考虑3D姿态和3D形状,允许进行综合评估和消融研究。通过与人类表现的比较,我们建议了几个未来的进一步改进。
Autonomous driving has attracted remarkable attention from both industry andacademia.An important task is to estimate 3D properties(e.g.translation,rotation and shape) of a moving or parked vehicle on the road.This task, whilecritical, is still under-researched in the computer vision community -partially owing to the lack of large scale and fully-annotated 3D car databasesuitable for autonomous driving research.In this paper, we contribute thefirst large-scale database suitable for 3D car instance understanding -ApolloCar3D.The dataset contains 5,277 driving images and over 60K carinstances, where each car is fitted with an industry-grade 3D CAD model withabsolute model size and semantically labelled keypoints.This dataset is above20 times larger than PASCAL3D+ and KITTI, the current state-of-the-art.Toenable efficient labelling in 3D, we build a pipeline by considering 2D-3Dkeypoint correspondences for a single instance and 3D relationship amongmultiple instances.Equipped with such dataset, we build various baselinealgorithms with the state-of-the-art deep convolutional neural networks.Specifically, we first segment each car with a pre-trained Mask R-CNN, and thenregress towards its 3D pose and shape based ona deformable 3D car model withor without using semantic keypoints.We show that using keypoints significantlyimproves fitting performance.Finally, we develop a new 3D metric jointlyconsidering 3D pose and 3D shape, allowing for comprehensive evaluation andablation study.By comparing with human performance we suggest several futuredirections for further improvements.人工智能论文:APOLLOCAR3D:一个大型3D汽车实例,了解自动驾驶的基准(ApolloCar3D: A Large 3D Car Instance Understanding Benchmark for  Autonomous Driving) al1L7lliF6lW8lfw.jpg
URL地址:https://arxiv.org/abs/1811.12222     ----pdf下载地址:https://arxiv.org/pdf/1811.12222    ----人工智能论文:APOLLOCAR3D:一个大型3D汽车实例,了解自动驾驶的基准(ApolloCar3D: A Large 3D Car Instance Understanding Benchmark for  Autonomous Driving)
回复

使用道具 举报

您需要登录后才可以回帖 登录 | 立即注册

本版积分规则 返回列表 发新帖

nic28当前离线
新手上路

查看:3228 | 回复:0

快速回复 返回顶部 返回列表