当前位置：网站首页>3D object detection dataset

3D object detection dataset

2022-08-02 05:23:00 【hello689】

Before on zhihu search to an article about data sets,太优秀了：Automated driving source data sets summary.Although there are many such data sets are introduced,But I still want to arrange、输出一下.After learning some knowledge,通过自己的理解,整理,再输出一遍,Help yourself remember a little.
在这里插入图片描述

image source: 文献 nuScenes: A multimodal dataset for autonomous driving

1. KITTI Dataset

数据集地址：The KITTI Dataset

论文地址：Vision meets robotics: The KITTI dataset

3D object detection benchmark 由:

7481张train image,
7518张test image,
And the composition of the corresponding point cloud,
共包含80256The goal of mark.
9个类别

评价标准：使用PASCALStandard to evaluate3D目标检测性能.对于汽车,我们要求3DBounding box overlap70%,And for pedestrians and cyclists,我们要求3DBounding box overlap50%.在2019年10月8日,KittiWebsite to followMapillaryTeam in their paper Disentangling Monocular 3D Object Detection中提出的建议,使用40A recall of position insteadPascal VOCBenchmark proposed11A recall of position.The comparison results is fair.The specific method for evaluating the data,There are many online blog has introduced,我就不再赘述：KITTI 3D目标检测的评估指标_W1995S的博客-CSDN博客.Understand the two concepts in advance is needed here：

mAP（mean Average Precision,All the class label of the average precision rate）
PR曲线

img source: The KITTI Vision Benchmark Suite (cvlibs.net)

Annotation data is introduced：

000000.txt:

Pedestrian 0.00 0 -0.20 712.40 143.00 810.73 307.92 1.89 0.48 1.20 1.84 1.47 8.41 0.01

1 类型：包括Car、Van、Pedestrian、Cyclist、Truck、DontCare等
2 截断：从0(未截断)到1(截断)浮点数,Truncation refers to the left image boundary object
3 遮挡：用整数 (0,1,2,3) To represent the state of several:
0 =完全可见, 1 = 部分遮挡
2 = A wide range of keep out, 3 = 未知
4 alpha： Observe the Angle of the object,范围[-pi , pi]
5-8 bbox： contains left, top, right, bottom pixel coordinates
9-11 Dimension shape dimension： 3D object dimensions: height, width, length (in meters)
12-14 位置：在相机坐标系下3D object location x,y,z(in meters)
15 rotation_y： In the camera coordinate system about abouty-axis的旋转,范围[-pi, pi]

注：在这里,“DontCare”Label for objects not marked area,Because they are too far from the laser scanner.

000000.png
在这里插入图片描述

000001.txt:

Truck 0.00 0 -1.57 599.41 156.40 629.75 189.25 2.85 2.63 12.34 0.47 1.49 69.44 -1.56
Car 0.00 0 1.85 387.63 181.54 423.81 203.12 1.67 1.87 3.69 -16.53 2.39 58.49 1.57
Cyclist 0.00 3 -1.65 676.60 163.95 688.98 193.93 1.86 0.60 2.02 4.59 1.32 45.84 -1.55
DontCare -1 -1 -10 503.89 169.71 590.61 190.13 -1 -1 -1 -1000 -1000 -1000 -10
DontCare -1 -1 -10 511.35 174.96 527.81 187.45 -1 -1 -1 -1000 -1000 -1000 -10
DontCare -1 -1 -10 532.37 176.35 542.68 185.27 -1 -1 -1 -1000 -1000 -1000 -10
DontCare -1 -1 -10 559.62 175.83 575.40 183.15 -1 -1 -1 -1000 -1000 -1000 -10

000001.png
在这里插入图片描述

Easy、Moderate、Hard标准定义如下：

Easy: Min. bounding box height: 40 Px, Max. occlusion level: Fully visible, Max. truncation: 15 %
Moderate: Min. bounding box height: 25 Px, Max. occlusion level: Partly occluded, Max. truncation: 30 %
Hard: Min. bounding box height: 25 Px, Max. occlusion level: Difficult to see, Max. truncation: 50 %

2. Waymo Open Dataset

数据集地址：https://waymo.com/open

论文地址：

Scalability in Perception for Autonomous Driving: Waymo Open Dataset（2020）
[Large Scale Interactive Motion Forecasting for Autonomous Driving :The WAYMO OPEN MOTION DATASET（2021）]([2104.10133] Large Scale Interactive Motion Forecasting for Autonomous Driving : The Waymo Open Motion Dataset (arxiv.org))

整个数据集包含1150个场景,Each time about20秒,且LiDAR和CameraIs a synchronous and calibration of. The image and laser radarbounding boxCarefully mark,And between each frame using a consistent identifier.

数据集介绍:

1150个场景,Each scenario length is about20秒
1000个训练集,150个测试集
1200万个LiDARComment box and1200M image comment box
Perception Dataset：标签为Vehicles, Pedestrians, Cyclists, Signs 四类
Motion Dataset：标签为Vehicles, Pedestrians, Cyclists三类.

3. NuScenes DataSet

在这里插入图片描述
image source： https://www.nuscenes.org/

数据集地址：https://www.nuscenes.org/download

论文地址：nuScenes: A Multimodal Dataset for Autonomous Driving

Source of data sets a brief introduction to：

MotionalHyundai motor group andAptivA joint venture of unmanned company,Committed to self-driving car safety,Reliable, and can reach.It has a unique power,Allow us to fundamentally change people's way of life.MotionalMission is to create and develop can save life、Technology save time and money,The works to change the world.

nuScenes数据集是由MotionalTeam development for unmanned public large data sets.In order to support the public research in computer vision and automated driving,Motional公开了nuScenes的部分数据.

关于nuScenes:

1个LiDAR,5个RADAR,6个camera,IMU,GPS
1000 scenes of 20s each
1,400,000 camera images
390,000 lidar sweeps
Two diverse cities: Boston and Singapore
1.4M 3D bounding boxes manually annotated for 23 object classes
New: 1.1B lidar points manually annotated for 32 classes
300GB

其中激光雷达：32线、范围为80-100m

在这里插入图片描述

在这里插入图片描述
image source: https://gas.graviti.cn/dataset/motional/nuScenes

Compared with other data sets：
在这里插入图片描述
image source: 论文 nuScenes: A multimodal dataset for autonomous driving

4. Appllo Scape

数据集地址：Apollo Scape

论文地址：The ApolloScape Dataset
在这里插入图片描述
image source: The ApolloScape Dataset for Autonomous Driving

ApolloScape由RGBVideo and the corresponding dense point cloud.包含140K张图片,And each image has the semantic information of pixel level.In the domestic data,So compared to some of the data set abroad,ApolloScapeThe data set contains traffic scene is very complicated,The first is the picture of one more,车多,Moving object more. 此外,This data set also includes lane based on color and style of different lane marker.
在这里插入图片描述
image source: The ApolloScape Dataset for Autonomous Driving

具体特点：

与kitti类似,也分为easy、moderate、hard三个子类.
标注信息包含28Class lane line
使用了2D/3DCombined with channel,是包含3DLabeling information of open data set

Compared with other data sets：
在这里插入图片描述
image source: The ApolloScape Dataset for Autonomous Driving

5. Lyft L5

数据集地址：

论文地址：

Self-driving Motion Prediction Dataset

Data set is introduced address：

This data set is divided intoPrediction Dataset和PerceptionDataset.

Prediction Dataset

The paper addresses of the data set：Self-driving Motion Prediction Dataset

超过1000Hours of autonomous movement data set.这是一个由20Self-driving car convoy of in 州Palo AltoA fixed route took on4Data collected in months.由17All of scene,Each scenario is long25秒, To capture the perception of the autopilot system output,Autopilot system coding traffic near、Bicycles and pedestrians precise location and move 作.除此之外,The data set contains a also contain15242A high-definition semantic markup elements and the area of hd An aerial view definition.总数据量有1000多个小时,接近90GB.

Data sets, there are three main parts：

17Thousands of scene,每25秒,Capture the self-driving cars and the movement of the traffic participants around it.
A high-definition semantic map,Capture the rules of the road、Track geometry and other transportation elements.
The region's high resolution aerial photographs,Can further help predict.

Perception Dataset

1.3M 3D annotations
30K lidar point clouds
350+ scenes at 60-90 minutes long

This data set contains training set test set and validation set,大小为120GB左右.This url can be at the bottom of the manual withQWEASDOr up and down keys to watch point clouds and the correspondingRGB图像数据集.
在这里插入图片描述
image source: Perception - Level 5 (level-5.global)

6. Argoverse

7. H3D

8. Cityscapes

The city landscape data set（Cityscapes）Mainly focus on the city streets2D语义分割,包含30个类别,Fine label5000张图,Coarse mark20 000张图片.2020年10月上线了3D bounding box的标注,Mark points and the results are divided into8大类下的30类标签,Simplified version labels contain19类.

1024x2048RGB图像,采样频率17Hz,16位深HDR;At the same time also provide compression8位深LDR版本.

9. CADC

地址：https://link.zhihu.com/?target=https%3A//arxiv.org/abs/2001.10117

The university of Toronto, Canada with unique advantage to provideSnow driving data set,Scenario highly targeted.

With the data type of too little,只有3D bounding box
Tag type is rough
Acquisition route fixed,Scene is relatively drab
The original data for500GB

10. Oxford RobotCar

All were very large data sets.Collect route isOxford中心,Because collected many times different weather and traffic condition data,The final dimension ridiculously large.但是,Data set provided only image、雷达、激光雷达和GPS/IMU的原始数据,And no additional annotations.