Build a defect detector

This tutorial walks the entire AcuSight MLOps loop with a concrete example: a visual inspection model that sorts cable glands into pass and fail. It’s the same flow as the video, but slower and with screenshots — a good first run if you’re new to the platform. By the end you’ll have a trained model deployed to a device and catching defects on a live video feed.

​

Before you start

You’ll need an AcuSight account and some images to work with — either an edge device that’s streaming images, or a folder of images on your computer. No local install required.

​

Step 1 — Create a project and bring in images

Open Projects and create one — call it Cable glands. A project is the workspace that ties your images, dataset, and models together. Then add images:

• From a device — images your devices capture arrive automatically as batches. Open a batch and choose Move Batch to Project. If you haven’t provisioned a device, see Connect a device.
• Upload — use Upload Data to add a folder straight from your computer.

Aim for a spread that reflects reality: mostly good parts, plus enough examples of each defect you care about. See Projects & batches for how images flow in and get organized.

​

Step 2 — Annotate

Open Annotate. Your batch starts under Unassigned — open it to begin labeling. Draw a box around each part and assign it a class: pass or fail. You manage classes and their colors on Classes & Tags. Mark each image as reviewed as you go. Consistent, tight boxes matter more than volume here — see Annotation for labeling guidance.

​

Step 3 — Freeze a dataset version

Once enough images are labeled, add them to the dataset and pick a train / validation / test split. Then open Snapshots → Create Snapshot to freeze an immutable dataset version — the exact set of images and labels your model will learn from. Optionally add:

• Preprocessing — for example, drop un-annotated images.
• Augmentation — for example, horizontal flips, to help the model generalize.

Snapshots are immutable, so every model traces back to the exact data it was trained on. More in Dataset versioning and Augmentation.

​

Step 4 — Train

Choose Train Model on your dataset version. Pick an architecture — YOLOv8 is a solid default — and a size. Nano trains fast and runs well on edge devices. Watch mAP, precision, and recall as they climb. For what the numbers mean and how to tune a run, see Training & experiments.

​

Step 5 — Evaluate and promote

Before deploying, check the model actually holds up. On the model’s page, review its metrics against the test split — the data it never trained on. When a model beats your current best, promote it to champion so it becomes the one that ships. See Model eval & champion promotion.

​

Step 6 — Deploy and watch it run

Open Models, select your model, and choose Deploy, then pick the group to deploy to. Every device in it picks up the model on its next check-in. Open the Video tab to watch live inference — detections appear in real time, with Good / Defect events listed alongside. That’s the full loop. See Deploy a model for rollout options like targeting a whole group of devices at once.

​

Keep it accurate over time

Parts change and new defects appear. Adapting the model is the same path you just walked: collect more images, re-annotate, snapshot, retrain, and redeploy. Because every model is tied to an immutable dataset version, you can always compare a new model against the one it’s replacing.

​

Next steps