Fig. 6: Conceptual illustration of the proposed Asynchronous Tracking and Context-Aware Sampling.
From: Bridging the latency gap with a continuous stream evaluation framework in event-driven perception
a Architecture of Asynchronous Tracking. A slow, high-fidelity base model (orange) performs full inference on event segments, generating initial bounding boxes and sharing features with a fast residual model (green). The residual model recursively updates predictions using shared features and new events, producing high-frequency outputs between base model cycles, leveraging temporal continuity of event stream to boost throughput. b Qualitative example of Context-Aware Sampling in sparse-event scenarios. Top row: Baseline model fails to localize the target (red box) as event density drops. Bottom row: Enhanced model detects sparse events, enters an inactive state, and reuses the last correct prediction (dashed green box) until dense events trigger accurate inference, preventing error accumulation. c Qualitative example of Context-Aware Sampling mitigating target drift during prolonged inactivity. Top row: Baseline tracker accumulates errors over time and loses the target. Bottom row: Enhanced tracker uses a timer to force reactivation after prolonged inactivity, re-localizing the slowly drifting target.
