Why This Matters

There is no single ideal place to run every computer-vision workload.

Some tasks benefit from running inside a camera. Others require an edge computer with a CPU, GPU or specialised accelerator. Some workloads may still benefit from cloud resources.

Our research explores how to decide where each part of a visual pipeline should execute.

The Computing Spectrum

A visual application can potentially use several computing layers:

Camera → Local edge device → Accelerated edge → Remote infrastructure

Each layer has different advantages.

Location	Typical strengths
Camera	Minimal bandwidth, immediate local processing
Edge CPU	Flexible local control and integration
Edge GPU / accelerator	Higher AI performance close to the source
Remote compute	Elastic resources and centralised heavy workloads

Research Questions

• Which workloads should remain closest to the camera?
• When does moving inference to an accelerator justify the additional hardware?
• How should a system operate when connectivity is limited?
• How can several types of hardware cooperate efficiently?
• How much data needs to leave the deployment site?
• How should workloads adapt when available compute changes?

Operational Trade-offs

The decision is rarely based on performance alone.

Important factors include:

• Latency.
• Network bandwidth.
• Energy consumption.
• Hardware cost.
• Resilience.
• Privacy requirements.
• Number of cameras.
• Model complexity.
• Maintenance requirements.

Heterogeneous Execution

Future deployments may contain different combinations of CPUs, GPUs, NPUs and camera-integrated processors.

A flexible vision architecture should be able to assign workloads to the most appropriate compute resource instead of assuming one hardware profile everywhere.

This research contributes directly to our goal of making advanced visual intelligence practical across very different real-world environments.