Products on Show
Synthetic sensor data platform for machine learning
Representatives from Applied Intuition will be on hand at the expo to share more about the company’s research on effective applications of synthetic training data.
Applied Intuition’s synthetic sensor data platform empowers machine learning (ML) and perception engineers to obtain the exact training data they need without the complexity or cost of data collection and labeling.
The quantity and quality of training data determine the performance of ML models used by autonomous systems. Unfortunately, data collection is hindered by operational costs and the difficulty of observing rare events, and data annotation is limited by the cost and accuracy of manual labeling. Some dense annotations, such as depth, semantic segmentation and optical flow, are impractical to obtain at scale. Collectively, the downsides of real labeled data reduce the velocity of autonomy programs, increase their overhead and limit the robustness of vehicles to edge cases.
With synthetic data sets, ML engineers simply specify the classes, environments and sensors they want to be represented in their data set and receive synthetic training data in return. Synthetic data is programmatically annotated, saving the cost, delay and inaccuracy of manual labeling.
At their core, synthetic data sets leverage the physics-based sensor and world models utilized by Applied Intuition’s sensor simulator, Spectral. This technology minimizes the domain gap between the synthetic data and the task domain by modeling the exact sensors the ML model is deployed on and the environment it will be deployed in.
In addition to the core functionality offered by Spectral, synthetic data sets automate the generation of diverse, realistic scenes to enable rapid data set definition. Smart behaviors enable actors such as cars and pedestrians to realistically interact with the world without any manual path specification. When required, default class, environment and event distributions can be overridden to provide exhaustive control over data set contents.