Cross-camera collaboration is emerging as a crucial strategy to support dynamic and heterogeneous real-world workloads in real-time video analytics (VA) systems. However, existing offloading methods lack the capability to jointly model spatial and temporal dynamics, which limits their offloading accuracy, leading to low system performances. We propose OctoCross, a spatiotemporal offloading framework for request offloading in multi-camera VA systems. OctoCross introduces a novel spatiotemporal model that jointly learns spatial and temporal dynamics to generate actionable offloading plans end to end. Extensive evaluations on real-world multi-camera datasets show that OctoCross achieves up to 5.8 \(\times \) higher throughput and 3.2 \(\times \) lower latency compared to baselines, while maintaining high SLO compliance across diverse deployment scenarios.