Parking Datasets | ML model training

Challenges

APS must function well across diverse parking operational design domains (ODDs), requiring high-performing perception models. However, training data with accurate labels and sufficient diversity can be hard to find. Collecting large quantities of real-world data can be slow, expensive, and even dangerous. Manual data labeling creates additional costs and delays between identifying a model failure and re-training with new data.

Why Parking Datasets?

Applied Intuition’s Parking Datasets empower ML engineers to develop robust perception models.

Bootstrap new ML models or complement existing training data with multi-modal, multi-sensor labeled training data for parking use cases. Address class imbalances by varying ODD elements and targeting rare classes.

Expand your parking ODD to new locations and train with dense, per-pixel labels that are difficult to obtain from real data, such as optical flow.

Benefits

Expand coverage

Expand an ODD to include elements not included in existing training datasets. Utilize synthetic data to address edge cases by targeting data sparsity issues or class imbalances.

Improve model performance

Train ML models on synthetic data to improve performance on long tail cases by up to 3x.

Iterate faster

Augment real data with synthetic datasets to reduce the cycle time and cost for collecting and labeling new training datasets when failures occur in testing and production.

Key components

Sensor models

Physically accurate synthetic data generated with validated, hardware-specific sensor models.

Ground truth labels

Error-free ground truth labels in industry-standard formats.

Minimize the domain gap

Mitigate the simulation-to-real domain gap with parameters to tune sensor and material behavior and domain adaptations to remove synthetic-specific features.

Environmental variants

Procedurally generated parking structures with thousands of variations in parking spot markings, materials, vehicles, VRUs, and other ODD elements.

Scene diversity

Directly define distributions over all scene components such as environments, weather, and lighting to maximize diversity and target specific edge cases.

Global coverage

Regions include the U.S., Canada, Europe, Japan, China, South Korea, and more.

Use cases

Perception and localization

Sensor suite design and validation

Automated parking development solution

Get started with Parking Datasets

Request a data sample and learn how Parking Datasets can accelerate your ML training for APS development.

ADAS and AD development platform

Vehicle software platform

Autonomy stack