Drivable area detection is arguably the most important task when building autonomous driving systems; it’s the core ingredient required to help autonomous vehicles understand safe driving areas and road conditions (detecting pedestrian crossings with bounding boxes, annotating road edges and lane markings with 3D polylines, and labeling arrows with polygons, for example).
TL;DR: Sama is thrilled to announce new and improved features for 3D point cloud annotation to provide our clients with the most secure, cost-effective, and fastest path to quality 3D drivable area annotation at scale.
Historically, labeling for this task has been approached with 2D vector and raster segmentation or 3D point level segmentation, which assigns semantic labels to every single pixel that belongs to an object of interest for a perception model. While the latter labeling type is highly precise, it requires a heavy lift in a multitude of ways.
Not only does it require a significant lift on the part of annotators to train, label, and QA the data, the resulting output requires an enormous amount of storage space and processing power.
Taken together, these elements can lead to a needlessly resource-intensive and expensive data annotation process.
A faster, more precise, and more cost-effective approach to data labeling for 3D drivable areas
Traditionally, annotation for 3D drivable areas (detecting lane marking, for example) might be solved by labeling in a 2D world and projecting into 3D. This approach, however, can come with significant consequences in regards to resulting quality: just 1 or 2 misplaced pixels in 2D can translate into meters of lost accuracy in the 3D world.
Others will approach the problem of labeling drivable areas using 3D segmentation, but there are many use cases in which the precision of 3D segmentation is not worth the time or monetary investment. For instance, you may only want to define a specific region of interest in a point cloud, and entire areas may not need to be annotated to adequately feed your model.
For these use cases, polylines and polygons can deliver comparable accuracy — but in significantly less time and requiring exponentially less storage space.
Polylines are an ordered collection of contiguous line segments that are created together as a single object, defining only areas of interest in the ground that absolutely need to be stored.
Sama has been producing 3D point cloud labeling for many of the top OEMs and Tier 1 suppliers in the world, providing them with drivable area data annotations quickly and cost-effectively without compromising on quality.
What’s more, Sama’s secure and compliant annotation platform and ISO-certified delivery centers protect our customers from costly mistakes that arise from poor security protocols. Coupled with our full-time dedicated workforce of annotators, you can rest assured that your data is protected from ingestion to delivery.