Edge AI

Edge Vision Pipelines without the Hype

Practical image pipelines from sensor capture through lightweight model deployment at the edge.

Format: Blended · Level: Advanced · Duration: 6 weeks

Tuition (informational): 1,450,000 KRW

Lead mentor

Rina Okada

Shipped on-device perception for warehouse robotics; cares about honest failure modes.

Overview

You build a pipeline that respects memory budgets and thermal envelopes. We stay away from benchmark theater; instead you profile frame drops and decide when to crop, quantize, or push work upstream. Labs pair a small camera module with a gateway so you can compare on-device versus near-edge routing.

What is included

Capture-to-tensor hygiene checklist
Quantization lab with side-by-side perceptual notes
Thermal throttling drill with scripted load
Gateway offload decision matrix
Dataset hygiene session with anonymized samples
Observability hooks for frame timing
Peer review rubric for model cards

Outcomes

Document a frame budget for a chosen SKU
Ship a quantized model with rollback metadata
Present a gateway offload plan to a mock cloud stand-up

Participant notes

Quantization lab was blunt in a good way—finally a class that admits perceptual loss instead of chasing leaderboard points.

Leo · 4/5
Gateway offload matrix is pinned above my desk. The Edge Vision Pipelines workbook references the exact telemetry fields we argued about in sprint review.

Aya Morimoto · Riverline Robotics · Google

FAQ

Will we train giant models?

No. You tune and deploy modest architectures suited to constrained devices. Large training runs are out of scope so we can keep feedback loops tight.

Can I use my own camera module?

What should I expect to struggle with?

Back to course table · Talk with moderators