The design, development and validation of in-cabin perception systems is a complex task. The variety of use cases, normative requirements in different geographies, the diversity of deep-learning models, the scarcity of data and accurate ground truth… Having all these into consideration, how can we take advantage of the latest advancements in generative AI to shorten the development lifecycle?
We propose a new approach using generative AI to create scene descriptions from false positive and false negative images and specific natural language prompts (VQA). We can then feed these descriptions to a synthetic data generation pipeline that will automatically create data and accurate ground truth to fill the data gaps.
We’ll see that we can use the same approach for in-cabin perception system validation. For example, to generate synthetic data for all validation cases from the EuroNCAP requirements using only natural language requirement descriptions from the documents.