Multi-agent Perception via Co-attentive Communication Mechanism

Multi-agent collaborative perception has the potential to significantly enhance perception performance by facilitating the exchange of complementary information among agents through communication. Effective communication plays a crucial role in enabling agents to collaborate and exchange valuable information. However, traditional methods face challenges when it comes to communication scheduling. To address these issues, we propose the Co-Attentive Multi-agent Perception (CAMP) model, a learning-based framework for multi-agent collaborative perception. In CAMP, we propose a novel co-attentive scheduler to construct communication among agents from both pixel-level and feature-level perspectives. We conduct extensive experiments on a public dataset and the experimental results clearly illustrate the effectiveness of our proposed CAMP in the multi-agent perception task while using the least amount of bandwidth resources.