Answers are grounded in this paper and cite their sources.
Resource-constrained automotive perception requires accurate and low-latency detection under limited onboard computing and energy budgets. Event cameras and spiking neural networks offer a promising solution through temporally sparse sensing and event-driven computation. However, strictly discrete spiking processing may limit adaptation to time-varying event activity and weaken the spatial information required for accurate localization. This study proposes LiquidSpike, a continuous–discrete hybrid framework that selectively introduces continuous computation at information-critical stages while retaining discrete transmission through most backbone and neck stages. LTC-IPLIF provides input-conditioned temporal integration, LiquidSSM performs localized long-range spatial state propagation followed by discrete re-encoding, and response-aware multi-scale processing improves cross-scale fusion and spatial refinement. On the Gen1 automotive event-camera dataset, LiquidSpike achieves 69.4% mAP@50 and 41.8% mAP@50:95 with 18.9 million parameters under the 4 × 2 configuration. Compared with SpikeYOLO under the same configuration, it improves the two metrics by 2.2 and 1.4 percentage points, respectively, while reducing the parameter count by 18.2%. Results on MS COCO 2017 provide supplementary evidence of applicability to temporally encoded frame-based inputs. These results demonstrate a competitive trade-off among detection accuracy, model size, and theoretical operation-level energy.
No takes yet. Share an insight, caveat, or question.
Yin et al. (2026) studied this question.
Synapse has enriched 10 related papers. Consider them for comparative context: