Learn / guide
Learn · Guides
Edge vs cloud computer vision — Australia buyer guide
Edge computer vision runs detection, tracking and event logic on-site — beside the cameras — so plants keep video local, cut latency and stay useful when the WAN drops. Cloud still has a role for training, batch analytics and shared management; the buyer question is what must stay at the edge versus what may leave the facility.
Pairs with computer vision →, AI hardware & edge → and manufacturing AI →.
Why Australian plants care about residency
- Customer and agency contracts often constrain where imagery and embeddings may live.
- Continuous multi-camera feeds are costly and risky to stream raw offshore.
- Retention, access and deletion are easier to prove when event video stays on-site by default.
- Private model and LLM choices for plant knowledge follow similar residency logic — see private LLM options in Australia →.
Latency and offline plants
Line-speed decisions and alerts cannot wait on a congested uplink. Edge inference keeps decode → detect → score → alert inside the plant network. When the WAN fails, recording and local alerting should continue; cloud sync can catch up with metadata and selected clips.
For remote sites with intermittent connectivity, design for store-and-forward — not “cloud or nothing”.
Edge-first event evidence vs cloud training / batch
A practical split many Australian industrial builds use:
- Edge — live decode, tracking, baseline comparison, severity scoring, local NVR, pre-/post-event clips.
- Cloud (controlled) — model training and evaluation, fleet dashboards, long-term analytics, approved clip upload for review.
- Not recommended by default — raw 24/7 video egress for “AI in the cloud” marketing claims.
See the illustrative path on visual OEE & edge event evidence → and the continuous monitoring shape on industrial camera AI Australia →.
When hybrid is the right answer
- Multi-site groups that need a central view but keep video and inference local per plant.
- Heavy model training on GPUs you do not want on the factory floor — then deploy weights to edge.
- Human review queues that pull only event clips into an approved AU region.
- Pairing vision events with plant knowledge assistants — residency still applies to prompts and docs.
Pitfalls
- Buying cloud-only vision because demos look slick — then discovering bandwidth and privacy blocks go-live.
- Ignoring NVR capacity, camera network segregation and edge appliance ops.
- Uploading “just a few” continuous streams that become permanent full egress.
- Treating embeddings and backup locations as free — they are still data residency.
Also read
Related guides & services
Need an edge-first vision architecture?
Bring your sites, cameras, WAN reality and residency constraints. We will recommend edge, hybrid or controlled cloud — then integrate with your ops tools.