Fig. 4: Impact of perception latency on event-driven trackers across datasets and hardware.
From: Bridging the latency gap with a continuous stream evaluation framework in event-driven perception
a Tracking accuracy (AUC) under STARE (-S, solid lines) vs. traditional frame-based evaluation (-F, dashed lines) on ESOT500-L. M_FETV2 stands for Mamba FETrackV2. Larger markers indicate faster inference. STARE reveals the impact of perception latency on accuracy hidden by traditional methods. Accuracy generally peaks at an optimal sampling window size under STARE. b Same as (a) but on ESOT500-H, validating STARE's consistency across dataset resolutions. The unimodal AUC vs. sampling window size trend persists. c STARE performance on FE10821: AUC vs. sampling window size for diverse trackers. The consistent unimodal trend across datasets (vs. ESOT500) highlights STARE's generalizability to external benchmarks. d STARE performance on VisEvent24: AUC vs. sampling window size. Results mirror (a-c), reinforcing the unimodal trend. e Ablation of sampling methods with HDETrack: Continuous Sampling (STARE, “Cont.'') vs. fixed-rate sampling from preprocessed frames ("Prep.''), using EventFrame (EF)21 and VoxelGrid (VG)21 representations. Continuous Sampling outperforms, leveraging event stream temporal continuity. Performance ranking reversals under STARE at 2 ms (f), 20 ms (g), and 50 ms (h) sampling window sizes. Faster models (e.g., DiMP50, MixFormer) outperform slower counterparts with high traditional accuracy, demonstrating STARE's bias toward throughput. i Accuracy degradation with simulated inference latency (speed multiplier < 1 slows inference). All models show monotonic AUC drops, quantifying latency’s direct impact. STARE performance on hardware with varying configurations: (j) RTX 3090 (high-power), (k) RTX 3080 Ti (lower-power, accuracy drop vs. 3090), (l) RTX 3080 Ti with parallel task contention (further degradation), illustrating how different hardware configurations impact on latency.
