数据集

数据集地址：The KITTI Dataset

页面解读

The 3D object detection benchmark consists of 7481 training images and 7518 test images as well as the corresponding point clouds, comprising a total of 80.256 labeled objects. For evaluation, we compute precision-recall curves. To rank the methods we compute average precision. We require that all methods use the same parameter set for all test pairs. Our development kit provides details about the data format as well as MATLAB / C++ utility functions for reading and writing the label files.
3D物体检测基准由7481张训练图像和7518张测试图像组成，并包含相应的点云数据，总共标注了80256个物体。为了进行评估，我们计算了精度-召回曲线。为了对不同的方法进行排名，我们计算了平均精度。我们要求所有方法在所有测试对中使用相同的参数集。我们的开发工具包提供了关于数据格式的详细信息，以及用于读取和写入标签文件的MATLAB/C++实用函数

• Download left color images of object data set (12 GB)
  下载物体数据集的左侧彩色图像
• Download right color images, if you want to use stereo information (12 GB)
  如果需要使用立体信息，下载右侧彩色图像
• Download the 3 temporally preceding frames (left color) (36 GB)
  下载前3帧时间序列的左侧彩色图像
• Download the 3 temporally preceding frames (right color) (36 GB)
  下载前3帧时间序列的右侧彩色图像
• Download Velodyne point clouds, if you want to use laser information (29 GB)
  如果需要使用激光信息，下载Velodyne点云数据
• Download camera calibration matrices of object data set (16 MB)
  下载物体数据集的相机校准矩阵
• Download training labels of object data set (5 MB)
  下载物体数据集的训练标签
• Download object development kit (1 MB) (including 3D object detection and bird’s eye view evaluation code)
  下载物体开发工具包，包括3D物体检测和俯视图评估代码
• Download pre-trained LSVM baseline models (5 MB) used in Joint 3D Estimation of Objects and Scene Layout (NIPS 2011). These models are referred to as LSVM-MDPM-sv (supervised version) and LSVM-MDPM-us (unsupervised version) in the tables below.
  下载预训练的LSVM基线模型，用于物体和场景布局的联合3D估计，这些模型在下表中被称为LSVM-MDPM-sv监督版和LSVM-MDPM-us无监督版
• Download reference detections (L-SVM) for training and test set (800 MB)
  下载训练集和测试集的参考检测
• Qianli Liao (NYU) has put together code to convert from KITTI to PASCAL VOC file format (documentation included, requires Emacs).
  Qianli Liao（纽约大学）编写了将KITTI格式转换为PASCAL VOC格式的代码，包含文档，需要使用Emacs
• Karl Rosaen (U.Mich) has released code to convert between KITTI, KITTI tracking, Pascal VOC, Udacity, CrowdAI and AUTTI formats.
  Kari Rosaen（密歇根大学）发布了代码，可以在KITTI、KITTI跟踪、Pascal VOC、Udacity、CrowdAI和AUTTI格式之间转换
• Jonas Heylen (TRACE vzw) has released pixel accurate instance segmentations for all 7481 training images.
  Jonas Heylen（TRACE vzw）为所有7481张训练图像发布了像素级别的实例分割数据
• We thank David Stutz and Bo Li for developing the 3D object detection benchmark.
  感谢David Stutz和Bo Li开发了3D物体检测基准
• Koray Koca (TUM) has released conversion scripts to export LIDAR data to Tensorflow records.
  Koray Koca（慕尼黑工业大学）发布了将LIDAR数据导出为TensorFlow记录的转换脚本

We evaluate 3D object detection performance using the PASCAL criteria also used for 2D object detection. Far objects are thus filtered based on their bounding box height in the image plane. As only objects also appearing on the image plane are labeled, objects in don’t car areas do not count as false positives. We note that the evaluation does not take care of ignoring detections that are not visible on the image plane — these detections might give rise to false positives. For cars we require an 3D bounding box overlap of 70%, while for pedestrians and cyclists we require a 3D bounding box overlap of 50%. Difficulties are defined as follows:
我们使用与2D物体检测相同的PASCAL标准来评估3D物体检测性能。因此，远处的物体会根据它们在图像平面上的边界框高度进行过滤，由于只有在图像平面上出现的物体才会被标注，因此Don’t Care区域中的物体不会被算作误报。需要注意的是，评估过程中不会忽略那些在图像平面上不可见的检测，这些检测结果可能会导致误报。对于汽车，我们要求3D边界框的重叠率达到70%，而对于行人和骑自行车者，我们要求3D边界框的重叠率达到50%。难度级别的定义如下：

• Easy: Min. bounding box height: 40 Px, Max. occlusion level: Fully visible, Max. truncation: 15 %
  简单：最小边界框高度：40像素，最大遮挡等级：完全可见，最大截断率：15%
• Moderate: Min. bounding box height: 25 Px, Max. occlusion level: Partly occluded, Max. truncation: 30 %
  中等：最小边界框高度：25像素，最大遮挡等级：部分遮挡，最大截断率：30%
• Hard: Min. bounding box height: 25 Px, Max. occlusion level: Difficult to see, Max. truncation: 50 %
  困难：最小边界框高度：25像素，最大遮挡等级：难以看见，最大截断率：50%
  All methods are ranked based on the moderately difficult results.

All methods are ranked based on the moderately difficult results.
所有方法都是根据中等难度的结果进行排名的

Note 2: On 08.10.2019, we have followed the suggestions of the Mapillary team in their paper Disentangling Monocular 3D Object Detection and use 40 recall positions instead of the 11 recall positions proposed in the original Pascal VOC benchmark. This results in a more fair comparison of the results, please check their paper. The last leaderboards right before this change can be found here: Object Detection Evaluation, 3D Object Detection Evaluation, Bird’s Eye View Evaluation.
注释2：在2019年10月08日，我们根据Mapillary团队在其论文Disentangling Monocular 3D Object Detection中的建议，将40个召回点替代原始Pascal VOC基准中提出的11个召回点。此更改使结果的比较更加公平，具体详情请参阅他们的论文。更改前的最后一次排行榜可以在以下位置查看：物体检测评估、3D物体检测评估、俯视图检测评估

Important Policy Update: As more and more non-published work and re-implementations of existing work is submitted to KITTI, we have established a new policy: from now on, only submissions with significant novelty that are leading to a peer-reviewed paper in a conference or journal are allowed. Minor modifications of existing algorithms or student research projects are not allowed. Such work must be evaluated on a split of the training set.
重要政策更新：由于越来越多为发表的作品以及现有工作重新实现的版本被提交至KITTI，我们制定了新的政策：从现在开始，只有具有显著创新性并且将用于发表会议或期刊论文的提交才被允许。对现有算法的微小修改或学生研究项目将不再允许提交，这类工作必须在训练集的划分上进行评估。

To ensure that our policy is adopted, new users must detail their status, describe their work and specify the targeted venue during registration. Furthermore, we will regularly delete all entries that are 6 months old but are still anonymous or do not have a paper associated with them. For conferences, 6 month is enough to determine if a paper has been accepted and to add the bibliography information. For longer review cycles, you need to resubmit your results.
为了确保该政策得到执行，新的用户在注册时必须详细说明其身份、描述工作内容，并明确提交目标会议或期刊。此外，我们将定期删除六个月以上且仍然匿名或没有关联论文的所有条目。对于会议论文，六个月足以确定论文是否被接受，并添加参考文献信息。如果您的论文审稿周期较长，您需要重新提交结果

Additional information used by the methods
方法使用的附加信息

• Stereo: Method uses left and right (stereo) images
  方法使用左图和右图（立体图像）
• Flow: Method uses optical flow (2 temporally adjacent images)
  方法使用光流（两个时间相邻的图像）
• Multiview: Method uses more than 2 temporally adjacent images
  方法使用超过两个时间相邻的图像
• Laser Points: Method uses point clouds from Velodyne laser scanner
  方法使用来自Velodyne激光扫描仪的点云数据
• Additional training data: Use of additional data sources for training (see details)
  使用额外的数据源进行训练（详见细节）

论文

论文地址：Vision meets robotics : The KITTI dataset

论文解读

摘要

We present a novel dataset captured from a VW station wagon for use in mobile robotics and autonomous driving research.In total, we recorded 6 hours of traffic scenarios at 10–100 Hz using a variety of sensor modalities such as high-resolution color and grayscale stereo cameras, a Velodyne 3D laser scanner and a high-precision GPS/IMU inertial navigation system. The scenarios are diverse, capturing real-world traffic situations, and range from freeways over rural areas to inner-city scenes with many static and dynamic objects.
我们提供了一个新的数据集，该数据通过一辆大众旅行车采集，旨在用于移动机器人和自动驾驶研究。总共记录了6小时的交通场景，采集频率为10-100HZ，使用了多种传感器，包括高分辨率色彩和灰度立体相机、Velodyne 3D激光扫描仪以及高精度GPS/IMU惯性导航系统。这些场景多样，涵盖了从高速公路、乡村地区到城市内部的实际交通情况，包含了大量静态和动态物体

Our data is calibrated, synchronized and timestamped, and we provide the rectified and raw image sequences. Our dataset also contains object labels in the form of 3D tracklets, and we provide online benchmarks for stereo, optical flow, object detection and other tasks. This paper describes our recording platform, the data format and the utilities that we provide。
我们的数据经过校准、同步和时间戳处理，并提供了校正后的图像序列以及原始图像序列。此外，数据集中还包含以3D轨迹形式的物体标签，并且我们提供了在线基准评测，用于立体视觉、光流、物体检测等任务。本文描述了我们的数据采集平台、数据格式以及我们提供的实用工具。

关键词

Dataset, autonomous driving, mobile robotics, field robotics, computer vision, cameras, laser, GPS, benchmarks, stereo, optical flow, SLAM, object detection, tracking, KITTI
数据集，自动驾驶，移动机器人呢，野外机器人，计算机视觉，相机，激光，GPS，基准测试，立体视觉，光流，SLAM，物体检测，跟踪，KITTI