Cost-Aware Scheduling Under Latency Constraints for Multi-View 3D Reconstruction Across the Edge-Cloud Continuum.

Area of Science:

• Computer Vision
• Machine Learning
• Distributed Systems

Background:

• Multi-view 3D reconstruction pipelines, particularly transformer-based ones, offer high accuracy.
• Deploying these pipelines across the edge-cloud continuum is difficult due to computational and data transfer demands.
• Balancing latency and cloud costs is crucial for effective pipeline scheduling.

Purpose of the Study:

• To develop a cost-aware scheduling strategy for multi-stage 3D reconstruction pipelines under latency constraints.
• To optimize the deployment of 3D reconstruction tasks across edge and cloud resources.
• To minimize cloud resource usage while ensuring real-time performance.

Main Methods:

• Implemented a service-oriented, multi-stage 3D reconstruction pipeline (depth estimation, multi-view fusion, point cloud merging).
• Experimentally characterized pipeline performance on diverse hardware, analyzing computational and communication latencies.
• Proposed an online scheduler with a top-K edge selection mechanism for dynamic stage placement.

Main Results:

• Identified significant imbalances between computation and communication latency due to large intermediate data.
• The proposed online scheduler effectively minimizes cloud costs while adhering to latency requirements.
• The top-K edge selection mechanism simplifies decision-making by considering network conditions and node utilization.

Conclusions:

• The developed scheduler significantly reduces cloud costs for 3D reconstruction tasks.
• The approach outperforms baseline strategies by dynamically optimizing edge-cloud resource allocation.
• This work provides an effective solution for deploying demanding 3D reconstruction pipelines in edge-cloud environments.