Towards Improved Perception System’s Generalization Through Generative Artificial Intelligence
Jiawei Wang, João R. Campos, Henrique Madeira · 2024
In recent years there have been significant advances in autonomous systems. Because they operate in complex and dynamic environments, these systems often integrate machine learning (ML)-based components, especially at the perception level. However, high-performance ML solutions require considerable amounts of data, and the resulting models are highly dependent on the training data. These models often leverage hidden patterns, known as dataset bias, that are not related to the concept that is being modeled, which, among other issues, limits the generalization and robustness of the models. Within the perception (images) domain, models trained on a specific dataset often fail to generalize to new datasets, despite the underlying concepts being the same. While various techniques have been proposed, such as diversity strategies, they fail to address the different biases in the data. This paper focuses on the issue of dataset biases for pedestrian classification, a core task for safety-critical autonomous systems. We evaluate the ML model’s generalization on four public reference datasets, along with the use of diversity strategies. To address the dataset biases, we then explore the use of generative artificial intelligence (GAI) techniques. The goal is to allow for the use of multiple large datasets to train a model for a given context, as well as tolerate drifts that might result from changes in the underlying system. Results show a noticeable improvement in the generalization performance of the models, highlighting the potential of such an approach.