The autonomous systems piloting self-driving cars will make decisions that have life-and-death consequences. There is little room for error in their ability to classify and track objects in a vehicle’s vicinity. That’s why in developing vehicle perception systems, curating high-quality training and validation datasets is critical. For example, seeing in training data how a pedestrian or cyclist responds over time to other vehicles or traffic signals in an urban setting can help a deep learning system understand the typical behavior of other pedestrians and cyclists on city streets, which in turn can inform a vehicle’s decision-making.
Recognizing the data challenges companies are grappling with in developing these systems, Navigant Research published a report that examines in depth several use cases, issues, and requirements for data used for autonomous driving and related mobility applications. Navigant senior research analyst Sam Abuelsamid authored the report, The Data Driving Automated Vehicles: Where it Comes From and Why Quality Matters, which Mighty AI commissioned.
You can download the full report here. In it, Navigant offers guidance on:
- When to optimize for real data and when to optimize for synthetic data
- Ways to achieve high-quality standards for your data labeling program
- Best practices for capturing real data from on-road driving
Key takeaways from Navigant’s research are outlined in the following graphic. Underscoring the critical role quality plays in curating datasets, Abuelsamid says, “Whether it is used for building, training, or validation, comprehensive and high quality datasets are essential. The output is only as good as the data fed into the system and low quality data can easily turn an otherwise viable product into something that may be worse than useless.”
Next month, Navigant will host a webinar during which Abuelsamid will moderate a discussion about the data driving autonomous vehicles with director of automotive for Nvidia Danny Shapiro, senior vice president for Toyota Research Institute Ryan Eustace, and Mighty AI’s own founder and CEO Daryn Nakhuda. Visit Navigant’s website to register for the webinar, happening on Tuesday, September 11, at 2 p.m. EST.
At Mighty AI, we’ve conquered the complexity of curating the highest-quality training and validation datasets that will help driverless cars understand their surroundings. Get in touch to learn more about how we can help your team.
Leave a Reply