In the field of machine learning and deep learning, datasets plays an irreplaceable role. No matter it is image data for convolutional neural networks or massive text data for natural language processing, the integrity and diversity of data directly determine the learning results of a model. With the advancement of technology, datasets that collected from specific scenarios have becomes the core strategy for improving model performance.

Annotation categories include: 3D point cloud, track ID, freespace (driving area, boundary line, segmentation), human body (bounding box, keypoint), vehicle (bounding box, 3D keypoint), lane line (lane line, edge lines), traffic signs (signs, lights), and human faces (keypoint, eyelid lines).

3D point cloud

3D point cloud annotation uses 3D boxes to label all movable objects in the radar map, such as cars, trucks, heavy vehicles, two-wheelers, and pedestrians.

Track ID

Vehicle track ID is to track the vehicle, pedestrian, two-wheeler in the picture. When labeling, it is necessary to ensure that the ID value of the same vehicle is kept consistent until the same ID disappears.

Freespace

● Drivable area

The drivable area is annotated with drivable area (the area that can be safely reached under the current state of the vehicle), invalid field of view, vehicle body, obstacles, parking rods, negligible area, pedestrians and deceleration zone.

● Ground sign segmentation

Ground sign segmentation encloses parking spaces, lane lines, arrows, zebra crossings, physical speed bumps, no parking areas, landmarks, etc., with lines of different colors to form a close-fitting polygon.

Human body

● Bounding box

It is divided into two frame attributes, normal frame (no occlusion and occlusion less than 80%) and ignore frame (less than the ruler and occlusion more than 80%).

● Keypoint

Keypoint is to add dot for pedestrians in the picture.

Vehicles

Vehicle in the picture is added frame, which is divided into three attributes: normal frame (no occlusion), imagination frame (partial occlusion), and ignore frame (smaller than the ruler).

Lane line

Distinguish the types of lane lines on the road and mark the corresponding attributes, or modify the incorrectly marked results (lane lines) in the figure.

Traffic signs

● Traffic signs plate

Mark all the traffic signs in the picture with frame (the traffic signs on the back are not marked). The attributes are mainly divided into three categories, namely effective frame (rectangle, circle, upper triangle, lower triangle, polygon), ignore frame and combo connection.

● Traffic light

Mark all the traffic lights in the picture with frame (the traffic lights on the back are not marked). The attributes are mainly divided into two categories: outer light frame (single light frame, multiple light frame, light off frame, ignore frame) and light wick frame (red light, yellow light, green light)

Face

● Annotate the faces (eyes, nose, mouth) in the picture with keypoint.

● Draw lines on the eyelids of the people in the frame.

End

Nexdata is the world’s leading AI data service provider. We have accumulated a large number of training datasets ( image, audio, video and text), and provide on-demand data collection and annotation services that power the world’s most innovative artificial intelligence.

If you need data services, please feel free to contact us: [email protected]

Facing with growing demand for data, companies and researchers need to constantly explore new data collection and annotation methods. AI technology can better cope with fast changing market demands only by continuously improving the quality of data. With the accelerated development of data-driven intelligent trends, we have reason to look forward to a more efficient, intelligent, and secure future.