Toward Object-Centric Learning for Perception in Autonomous Driving

Our work addresses the domain of road user perception in autonomous driving with minimal supervision. The first goal was to develop a flexible, end-to-end data-processing architecture adoptable to autonomous driving data of different sources for seamless integration into slot-attention based object-centric learning (SA-OCL) implementations. Next, we present results of transferring the capabilities of minimal-supervised SA-OCL to autonomous driving. The research was motivated by the importance of bridging the gap between synthetic and generic demonstrations of SA-OCL, and domain-specific targeted detection incorporating real-world complexities. We implemented a novel dataset generation workflow and validation framework that together provide the groundwork for determining regions of relevant road users from its background elements in diverse driving environments using a combination of feature and motion cues within a slot-assignment scheme. Our findings suggested that the proposed approach performs best for city and urban driving scenarios, in which road user activity is typically proportional to the proximity of the ego-vehicle, confirmed by higher resulting mIoU values for all types of road users as their detection area increases. We also provide a proof-of-concept, based on leveraging depth signals to guide future work of model scaling, and provide specific options for architectural improvements for robust generalization.