This repository provides synthetic, diverse, and realistic scenario log samples illustrating how to augment video for safety critical and rare scenarios in autonomous vehicular systems and smart city applications.
As featured in the NVIDIA Cosmos Cookbook's end-to-end recipe for Synthetic Data Generation (SDG) for Smart Cities, this repository serves as a foundational guide for obtaining sample CARLA scenario logs as input to synthetic data generation pipeline. These log files define complex, realistic movements of agents (vehicles and pedestrians) in a simulated environment.
In the context of the Cosmos SDG recipe, the scenario logs demonstrated within this repository are used to:
- Provide Ground Truth: Supply the initial high quality, realistic, reactive, and diverse synthetic data for any complex, rare, or safety critical scenario required for Stage 1: Generating Ground Truth with the CARLA simulator.
- Accelerate Sim-to-Real Transfer: Ensure the generated synthetic videos contain realistic traffic interactions and incidents, which are then used with the Cosmos-Transfer model to create photorealistic augmentations, ultimately bridging the domain gap for fine-tuning Vision-Language Models (VLMs) or other perception models used in Intelligent Transportation Systems (ITS).
- Enable Scalability: Logs generated with Inverted AI’s expertise are produced to the exact specification of the rare or safety critical scenario of interest, enabling easy access to data that is otherwise prohibitively expensive to capture or manually craft.
This repository is a sample of sophisticated traffic log data to train cutting-edge perception and AI models for urban environments.
See the Inverted AI blog post 'AI Behind the Wheel: How Inverted AI’s Scenarios Enrich NVIDIA’s Traffic Data Recipe' to learn more about generating synthetic data for traffic and visit the Inverted AI website to learn more about Inverted AI synthetic logs.