Products on Show
DAY 1: Synthetic sensor data platform provides high-fidelity training data in real time
Applied Intuition is presenting its synthetic sensor data platform that reduces the complexity and cost of data collection and manual labeling on Day 1 in Stuttgart. Its synthetic data sets enable ML engineers to immediately define and generate labeled data when they discover a shortcoming in their model, solving problems in days.
Training data is usually collected to address a specific issue identified in model analysis. As most real data is annotated manually, if the issue is due to a rare but important occurrence, it can cost dollars per frame for annotations such as semantic segmentation and take over a month for large batches of data. Human errors also degrade model performance.
Other annotations, such as optical flow for camera images, are created using another sensor modality such as lidar to obtain additional information when collecting the raw data. However, this approach is vulnerable to sensor calibration issues and is limited by the density of data produced by the additional sensor, which, in the case of lidar, is significantly less dense than a camera image.
With synthetic data sets, annotations are dense and error-free because they are generated programmatically with exact knowledge of the simulated world. Data sets with millions of examples of an edge case can be generated without spending years collecting examples. Additionally, randomization can be used to discover edge cases not yet observed in the real world, which can then be addressed before on-road failures occur.
"Production synthetic data set technologies enable autonomous vehicle perception and localization developers to generate, in real time, the specific labeled data that they need,” explained Chris Gundling, head of sensor simulation at Applied Intuition.
“Synthetic data can be generated at scale, without the need for human annotators, and can touch on conditions that are rare but critical to real vehicles on the road. The latest technology in rendering, GANs and sensor physics modeling has increased the fidelity of synthetic data to a point where significant gains in training and a minimal, manageable domain gap in inference are achieved," Gundling concluded.