OutStanding DataTech-Autonomous Driving

Autonomous Driving

Service Introduction

Data Type	Typical Dimension Objects	Labeling Method
2D	Obstacle	2D Box/Cuboid 3D Box
	Lane line/road edge/stop line callout	Line segment
	Parking space inner corner mark	Point
	Semantic segmentation	Polygon
	Signal Lights/Signs Callout	2D Box
3D	3D dynamic and static obstacle annotation	3D Box
3D	3D parking space callout	2D Box
2D & 3D	Fusion labeling	3D Box/3D Box Mapping
2D & 3D	Joint Callout	3D Box/3D Box Mapping

ADAS Scenarios

dimension	Category
Dataset Data Source	3D point cloud/2D GRBD/Image/3D point cloud and image combination
Data Volume	Large/Medium/Small Scale
Data Usage	High-precision map/automatic driving/parking/automatic parking
point cloud data grid	pcd/ply/bin/xyz/obj
data accuracy	2D image resolution (2k/4k) Number of channels 3D Point Cloud Harness (16/32/64)
Point Cloud Generation Method	RGB/Lidar
light conditions	bright/dim
weather conditions	sunny/cloudy/rainy/snowy/foggy
Labeling Tasks	(2D/3D) object detection/semantic segmentation
Collecting Data Driving Scenarios	city/town/high-speed/complex road conditions
Road markings-Traffic signs-Specification	Domestic/Europe and America
Dimension Type	Point Cloud/Image/Fusion Annotation manual marking/semi-automatic marking/automatic marking
Labeling Method	2D Rectangular Box/3D Box/Polygon/Line/Point
Continuous Frame Callout	trace point continuous dimension
3D point cloud annotation range	entire scene annotation/front camera visible range annotation only (refer to kitti)
Labeling Quality Control	full inspection/random inspection (random inspection by statistical category)
Dimension Tool	self-research tools/custom tools/customer tools/open source tools

Low-speed automatic parking scene

Parking space collection

Parking space collection usually uses a fisheye camera, and after the collection, the video is converted into images by extracting frames, and then the internal corner points of the parking space are annotated.

Overseas Collection

Overseas customers mostly use the reconstructed open world parking lot plane data collected by lidar to mark parking spaces in the point cloud.

Case Study: Annotation of situation and necessity of static obstacles

Static object position and object type	In front of the fence and RB annotation	Between fence and RB annotation	Behind fence annotation (RB annotation is not present)	Behind RB annotation
Reflective objects (e.g. TS, HRM)	Yes	Yes	YES but only in 5 meters distance from Road edge	YES but only in 5 meters distance from Guardrail
Non reflective (e.g. UBP, bridge pillar, highway accessories)	Yes	Yes	Yes but only in close distance from Road edge (~30cm)	No