Lidar Annotation with Rapid

Rapid currently supports 3D Lidar cuboid annotations. However the setup is a little bit different from a typical Rapid project.

This guide will walk you through setting up your first Lidar project and getting high-quality annotations back quickly.

Pipelines

Rapid currently supports

Standard pipeline

The standard pipeline is more useful for scenes with fewer frames in each task and not much cuboid movement across frames. For simpler, more static scenes, the quality results of the standard pipeline will be very similar to that of the dynamic pipeline.

Dynamic pipeline

The standard pipeline is more useful for scenes with many frames per task and larger number of cuboid paths that need to be labeled. The results of this pipeline will typically have better precision around each cuboid with the trade-off of taking longer due to increased complexity and being slightly more expensive. Currently, the default is the standard pipeline and you will have the option to select the dynamic pipeline after creating a project. You will have the opportunity to switch back and forth across pipelines until you have finalized your first batch.

Data Ingestion

Data Conversion

Raw Lidar annotation data must be converted into Scale's format before taskers can work on it. First, download this Github repository and follow the set-up instructions. Also be sure to read the Scale AI Lidar Toolkit for more detailed information. Examples of the expected data formats can be found in the Scale format folder of the repository. The example script can be used to process your data and upload it to your own cloud storage service, but be sure to remember to populate the bucket and path parameters with your own information. Also, note that the final line in that file to create task is not needed to upload data to S3.

Batch Creation

In Rapid, tasks are be associated to a batch. The next step after data conversion is to create a batch directly using the API. You can run the command in browser if you put your API key in the Authentication text field.

Create Lidar Annotation

Next, you can create a Lidar annotation task using this API endpoint attached to the batch you created. Feel free to read over the other tabs in the section for more information. Please note that Scale must be able to access the attachments you are using to submit the task. Check this document for information on setting up the permissions.

Batch Finalization

The last step is to finalize your batch using the API. After this step, taskers will be able to work on your project.

Quality

Auditing Tasks

Audits (discussed in the calibration batches page) are extremely important for maintaining quality across the board. Here, you will have the option to take a look at the tasker response and make changes as you see fit using Scale's lidar labeling tool.

Audit tool in action

Audit tool in action

Adjust cuboid dimensions in the side bar (left side of the audit tool)

Adjust cuboid dimensions in the side bar (left side of the audit tool)

With the camera intrinsics/extrinsics, we can project the cuboids drawn in 3D into the 2D space

With the camera intrinsics/extrinsics, we can project the cuboids drawn in 3D into the 2D space

Creating Quality Tasks

When you are satisfied with the task, you can also create quality tasks that will be served to train and evaluate taskers on your project using the "Create Quality Task" button on the bottom right.

Editing Quality Tasks

Sometimes quality tasks may not capture exactly what you want taskers to pay attention to. You can edit your project's quality tasks by simply clicking the link in the training / evaluation tasks page, which will bring up a similar Lidar editor as the audit tool.

Training tasks in the Quality Lab

Training tasks in the Quality Lab

Updated 24 days ago