CIOFL: Collaborative Inference-Based Online Federated Learning for UAV Object Detection

Federated learning (FL) has great potential in visual applications such as object detection by unmanned aerial vehicles (UAVs), since different UAVs can capture diverse characteristics of targeted objects from different angles. More importantly, how to collaboratively train an accurate object detection model via FL in an online manner is critical to UAV object detection in practice. In this demonstration, we present a working prototype of CIOFL on embedded computers, Collaborative Inference-based Online Federated Learning for object detection within a UAV swarm. In essence, the proposed CIOFL enables online FL for object detection among multiple nodes, by continuously adding high-quality real-world samples inferred just by these nodes with a complex and large-scale object detection model. Our evaluation results show that, compared to the traditional FL, CIOFL improves the convergence rate and accuracy by ~1.3✗ and ~ 1.34✗, respectively. We envision that, the CIOFL can effectively enhance the applicability of UAVs conducting object detection in practice.