vISP-in-the-loop: developing better perception stacks through cloud-to-car environmental parity
22 May 2025
Room A (W2/3)
Developments in scenarios, simulation, validation and in-the-loop testing (continued)
Increasing levels of driving automation in more complex ODDs demand effective ways to continuously develop and verify software across millions of virtual miles and thousands of driving scenarios. As a result, scalable MLOps and DevOps have become the engineering foundations underpinning those at the leading edge of AD technology. In this presentation, Arm and AWS join forces to discuss how cloud-to-car environmental parity can enable better testing coverage and faster time-to-market across the industry. Based on a practical SOAFEE-based case study, the presenters will explore the impact of adding the image signal processor (ISP) model into the virtual perception pipeline, and how a similar methodology can benefit OEMs, Tier 1 suppliers and AD software vendors.
- The pros and cons of different virtual ECU abstraction levels
- The advantages of cloud-to-car environmental parity for software development and verification
- Why images tuned for human vision may not yield best results for compute vision
- The importance of integrating a virtual model of the ISP into both AI training and testing stages