Fig. 7: Quantitative evaluation of model enhancement strategies under STARE.
From: Bridging the latency gap with a continuous stream evaluation framework in event-driven perception
a Accuracy (AUC) of OSTrack-based85 variants on ESOT500-L across sampling window sizes. Curves compare: Baseline (green), +Predictive Motion Extrapolation (+Pred, orange), +Context-Aware Sampling (+C, light blue), +Asynchronous Tracking (trained on 500 Hz annotations, yellow), and +Asynchronous Tracking (trained on 20 Hz annotations, gray). Asynchronous Tracking (500 Hz) combined with Context-Aware Sampling (+Async+C, dark blue) consistently outperforms other strategies. b Motion dynamism vs. Asynchronous Tracking effectiveness. Blue solid line: high-dynamism (left y-axis, negative performance gain from Predictive Motion Extrapolation, also defined as the unpredictability score) for ESOT500-L sequences. Orange dashed line: Accuracy improvement (right y-axis) from Asynchronous Tracking. Higher dynamism poses greater challenge. c Context-Aware Sampling performance in some sparse-event scenarios. Blue line: Baseline accuracy (left y-axis). Orange line: Accuracy with Context-Aware Sampling (left y-axis). Green dashed line: Sparsity Rate (right y-axis, percentage of model inactivity). Context-Aware Sampling demonstrates robustness in low-motion contexts.
