Josip Štajdohar

Learning-based multi-view 3D reconstruction pipelines, such as transformer-based approaches, enable the accurate reconstruction of 3D scenes from multiple images, but their deployment across the edge–cloud continuum is challenging due to high computational demands and large intermediate data transfers. Effective pipeline scheduling in the continuum must therefore balance latency constraints with the cost of cloud resource usage. In this work, we address cost-aware scheduling under latency constraints for a multi-stage 3D reconstruction pipeline consisting of depth estimation, transformer-based multi-view fusion, and point cloud merging with export to a rendering-ready representation. We implement a service-oriented pipeline where each stage can be executed either on edge or cloud nodes, and we experimentally characterize its performance on representative hardware platforms. The results show a strong imbalance between the computational time and communication latency across platforms, mainly due to large intermediate data. Based on these insights, we propose an online scheduler that dynamically selects stage placements to minimize the cloud cost while satisfying latency constraints. The scheduler incorporates a top-K edge selection mechanism that reduces the decision complexity by jointly considering the network conditions and node utilization. Simulation results parameterized with real-system measurements show that the proposed approach effectively reduces cloud usage while meeting latency constraints, outperforming the baseline strategies based on single-node pipeline execution.