Continuous Frame 3D Point Cloud Annotation
The Difference Between Image Data and Point Cloud Data
First, the perspective is different. Image data is a two-dimensional representation of the real world, while a Point Cloud contains three-dimensional information( a coordinate system), which can be projected to a variety of views.
Second, the data structure is different. Image data is regular, dense, and in order, while point cloud data is irregular, disordered, and sparse.
In terms of spatial resolution, image data is also much higher than point cloud data.
Image Data Disadvantages
However, the disadvantage of image data is that it is greatly affected by external lighting conditions, and it is difficult to apply to all weather conditions. LiDAR is not limited by darkness or extreme weather.
To a certain extent, it makes up for the shortcomings of cameras, and it can accurately perceive the distance of objects.
Now we’d like to share a 3D Point Cloud annotation case.
Point Cloud Data Annotation
Annotation Types:
3D Bounding box — A cuboid consisting of center point coordinates, length, width, and heading angles, which is used to locate the object’s position in the point cloud.
Obstacles. In the driving environment, all objects on the category list (under certain conditions) need to be labeled.
For example, mainly participants on the road (including pedestrians, vehicles, etc.) that are moving and stopping, and basic transportation facilities(including road markers and traffic signs).
Target ID
The target IDs are labeled according to the frames before and after. An object has only one target ID, and it cannot be reused.
There are 50–100 frames in a sequence frame, and the ID and size of the object within a sequence frame should be the same.
Three-dimensional Annotation Guidelines
(1) The bottom of the 3D bounding boxes of all dynamic obstacles on the ground should fit the ground, not higher than or lower (if there is a slope, the lowest point of the obstacle on the ground shall be considered as the measurement).
(2) The labeled obstacles shall not have missing points.
(3) 3D bounding boxes should fit the obstacles, that is, the obstacle should be marked with 3D bounding box with precision.
(4) Label the forward direction of vehicles, pedestrians, or people riding bikes or motorcycles.
(5) The rearview mirror, antenna, opened door, and trunk of the vehicle should all be labeled.
(6) Minimum annotated size: Objects with points greater than or equal to 10 points need to be annotated, those with less than 10 points do not need to be annotated.
(7) Tracking ID: the ID of the same target in the frames before and after must be consistent.
(8) Only the objects with the position and shape that can be determined should be annotated, and the category of obstacles that cannot be determined by 2D images and 3D point clouds should be annotated as others.
(9) Labeling range: Label the objects within the 150-meter-long radius of the ego vehicle.
Vehicle Annotation
1. Car
Cars with containers/shelves on the top shall be labeled, and when there are goods on the top of the car, the vehicle shall be labeled together with the goods.
Check the completeness of the point cloud. If the point cloud is dense and the actual size of the point cloud exceeds the normal size, use the actual size as the measurement. The height here depends on the specific situation, please make sure that it does not exceed the ground and it includes all the point clouds. When labeling cars, try to attach the border of the box to the object precisely, and ignore scattered points outside.
2. Large- and Medium-sized Bus
If there is a bus with two or three wagons running straight ahead, label the whole bus; When the vehicle is making a turn, if the front and rear parts are too large to be marked into one box, the bus body should be divided into two parts, and the front and rear parts should be labeled independently. (When the vehicle is straight ahead, the ID of the frame before and after shall be consistent. When the vehicle is making a turn, split the original box into two new and independent boxes at the turning point, and associate the two new boxes with the ID of the vehicle running straight)
3. Truck
(1) When a non-closed truck carries goods exceeding the size of the vehicle body, the goods shall be labeled together.
(2) When there is a car on the trailer, the car being carried should be labeled separately.
(3) The car on the vehicle carrier and the vehicle itself are labeled as one target.
4. Engineering Vehicle
(1) The forklift truck should be labeled together with the fork. When the goods are too large, only mark the forklift truck itself, and the direction of the fork is the front direction.
(2) For the very extreme and very long structure of the crane boom, only the main part of the vehicle that affects traffic is labeled, and the extended lifting part does not need to label.
5. Tricycle
(1) Tricycles ridden or pushed by people and their users are labeled as tricycles as a whole, and when people are separated from tricycles, they are labeled separately.
(2) If the tricycle carries objects that exceed the size of the vehicle, the goods will also be labeled as the tricycle.
The annotation of trucks (wagons, trucks, trailers, pickup trucks, etc.), buses (large buses, buses, school buses, etc.), and engineering vehicles (tank cars, cranes, forklifts, etc.) should be labeled according to the actual size.
Pedestrian Annotation
The specific requirements for pedestrian annotation are as follows:
(1) Pedestrians do not need to be divided into adults and children, and the same case for various postures (squatting people also need to be labeled as pedestrians).
(2) Pedestrians in the crowd shall be labeled separately. If they are too dense to separate a certain pedestrian outline, please refer to the frames before and after (if it is still hard to split, label them as a whole, and their ID should be a new ID).
(3) For the person holding objects, if it is hard to separate the object and the pedestrian and the pedestrian is small, the holding object will be marked with the pedestrian together.
(4) For people who are pulling suitcases or other objects, only box pedestrians themselves.
(5) Strollers with babies and wheelchairs with people should be marked with the user as a pedestrian. When there are no people setting in the stroller or wheelchair, they don’t need to be labeled.
(6) The box of the pedestrian should be adjusted according to the actual size.
(7) When the adult holds a baby, the baby and the adult are labeled in one box.
(8) For the adult and a child hand in hand, the child and the adult are labeled with two boxes separately.
In pedestrian annotation, there are four key points:
(1) The lower edge of the box must fit into the ground.
(2) The top edge of the box fits the top of the head.
(3) Shared size: if the human body does not contain outstretched arms and legs, only the body is labeled.
(4) The length and width of the pedestrian bounding box should be within 1m x 1m.
Bicycle and Motorcycle Annotation
(1) If a person is sitting or standing on a cycle, the cycle and its user need to be marked together, and the person standing by also needs to be marked together; In the labeling process, the cyclist in front generally has only in the back half of the point cloud. When labeling the cyclist in front, the right side of the box should match the outline of the point cloud. Similarly, the cyclist in the back has only the front half of the point cloud, so the left side of the box should match the outline of the point cloud.
Bicycles and motorcycles shall be labeled separately as much as possible. If the outline of a single bicycle or motorcycle cannot be separated, workers should take the frames before and after as reference (if it is still hard to split, it should be labeled as a whole, and the ID is a new ID).
The Annotation of Other Dynamic Obstacles
In addition to the dynamic obstacles on the road, the dynamic obstacles on both sides of the road also need to be labeled. When annotating, for the occluded objects, labelers need to image them based on the actual size.
The Specification of Obstacle Forward Direction
Obstacles such as vehicles, pedestrians, and bicycles need to be labeled with a forward direction. In the visualization dashboard, the direction of the 3D box should be consistent with the real direction.
The Specification of Box Size
(1) When the point cloud is dense, the size and the imaginary part should be adjusted according to the actual size.
Note: No point missing. Pay attention to the accuracy.
(2) When the point cloud is sparse, it should be labeled the same size as in the reference. If the car’s outline and size can be seen in the sparse point cloud, it should match the actual point cloud precisely.
Note: No point missing. Pay attention to the accuracy.
End
Outsource your data labeling tasks to ByteBridge, you can get high-quality ML training datasets cheaper and faster!
- Free Trial Without Credit Card: you can get your sample result in a fast turnaround, check the output, and give feedback directly to our project manager.
- 100% Human Validated
- Transparent & Standard Pricing: clear pricing is available(labor cost included)
