Sama was recently at the Auto.AI show in San Francisco and over half of the questions asked there were about LiDAR, point cloud data and 3D imagery. Below, we give you the view from our window.
What is LiDAR and Why is it Such a Hot Topic?
To answer the first: LiDAR (Light Detection and Ranging) is a laser-based surveying method that builds up a depth-based image of the world by shining out laser lights and then measuring how long it takes for the reflected pulse to bounce back to the sensor. (Read more about LiDAR technology and its uses in this Techcrunch article.)
To answer the second question; LiDAR is one of the critical technologies in the development and deployment of autonomous vehicles. It is also a part of the development mix in other hot topics like border security, drone mapping and much more.
Because of the ability to collect three-dimensional measurements, laser scanning systems are used for surveying the built environment (such as buildings, road networks, and railways) as well as creating digital terrain (DTM) and elevation models (DEMs) of specific landscapes.
Environmental applications also benefit from LiDAR – laser scanning is a popular method of map design, including mapping flood risk, carbon stocks in forestry and monitoring coastal erosion.
In the field of autonomous driving, LiDAR's strength is being able to visualize a 360 view around the vehicle. The only significant company currently not on board is Tesla, who is using camera and radar only at this point. Their abstention from the technology is also widely talked about. And LiDAR is the technology that Waymo/Google and Uber had gone to court over.
Key Challenge and Aspirations Around LiDAR and Point Cloud Annotation
Many companies - Sama included - can annotate 2D images, and recognize that LiDAR and 3D (also called Point Cloud) labeling is rapidly growing need. However, labeling 3D data presents some unique challenges.
Point Cloud Data labeling challenge 1 - navigating/labeling in a 3D space requires a carefully designed UI. Many companies develop proprietary tools - but then need to find a workforce that can be trained to use it. (Sama can help with that.) We are also in the process of developing our own Point Cloud Annotation tool, including a Point Cloud Viewer, that we can make available to clients.
Point Cloud Data labeling challenge 2 - depending on the sensor you are using, the resolution and the clarity can be miserable - making it hard to differentiate between objects. To get higher resolution, one solution is to use a more sensitive sensor. The top of the line model of these in existence today can cost between $60,000 to $80,000. These systems are primarily used in test and data collection vehicles and are too expensive for production cars.
The most common solution to this last challenge is that many companies use multiple, but less expensive LiDAR sensors. (We often see test cars from the major brands driving around our neighborhood in San Francisco. They’re usually set up with multiple sensors to build datasets for testing.)
From the conversations we had at Auto.AI, we learned that there is a growing need for a broadly available 3D annotation platform that has:
Labeling UI that is intuitive enough for a non-data scientist to use.
Flexibility to be used for more than one customer/industry
Access to skilled workers who can be trained to provide high quality and consistent annotation (did we mention we help with that?)
The ability to work on and transfer large files. Large 3D file sizes can make it challenging to share files over the cloud.
The ability to work on and transfer sensitive files. There is a growing need to ensure the security and privacy of data gathered.
The ability to produce a top quality data set - the basis of a top quality algorithm.
Trends in LiDAR Adoption
As we have come to expect with technology, investments lead to scale, which leads to price drops. In the past year, GM's Cruise project and Google spinoff Waymo, among many others, made significant investments in LiDAR technology, which is predicted to lead to an order of magnitude cost reduction.
Along with price drops, we also see considerable gains in capabilities. Tech leader Velodyne's new VLS-128 sensor set a new record by doubling the number of laser beams on its previous top-of-the-line LiDAR system to a massive 128 while shrinking the overall size of the sensor by 70 percent. Velodyne has announced price cuts on other products as well.
One trend that Sama can speak to personally is LiDAR equipment manufacturers partnering with annotation and labeling partners to develop a full, affordable solution to companies wishing to put this technology to use on their projects. (We help with that.)
We learned a lot talking to people at the AutoAI show and we look forward to being part of another lively conversation at the Nvidia GPU show later this month. (If you're going, can we buy you a drink?)
Please drop us a line if you would like to schedule some time to talk about your plans and aspirations in LiDAR or other data-related areas.
Steve is a Senior Account Executive for Sama focusing on AI applications for the automotive industry.