Data & datasets
curated workspace
The workspace where model-building happens. A project owns the batches assigned to it, its annotations, its dataset versions, its training jobs, and its models. Projects are scoped to an organization.
group of images
A collection of images that move through the pipeline together. Every batch carries two independent states: a status (
collecting → complete, tracking edge-upload progress) and a stage (unassigned → annotating → dataset, tracking its position in the project workflow).unassigned data
Where images land before they belong to any project — a batch with no project assigned. You browse it as a device’s Edge Data; once a batch is added to a project it shows up under Unassigned. Batches collect images until they fill up or go idle, then auto-complete and are ready to assign.
label on an image
A label applied to an image — for object detection, a bounding box with a class. Annotating happens in the browser and moves a batch into the
annotating stage.label category
A category an annotation can belong to (for example,
pass-blue or fail-white-nut), each with its own color. A project’s classes define what its models can detect; you manage them — and free-form tags — on the Classes & Tags page.immutable snapshot
A frozen, point-in-time snapshot of a project’s labeled batches, with train / validation / test splits. Versions are immutable — once created, their contents never change, which is what makes a training run reproducible. New labels or batches mean a new version. Also called a snapshot.
Models & training
unit of background work
An asynchronous task run by the ML service — most commonly
training, but also export, inference, and deployment. You watch a job’s progress and metrics as it runs.trained artifact
The output of a training job — weights plus metadata — registered against the project (and dataset version) that trained it. A model can be exported to several formats (PyTorch, ONNX, TFLite, and Synaptics NPU for edge hardware).
action
Sending a trained model to a group of devices. Every device in the group picks up the model on its next check-in, and its detection labels come from the dataset’s classes.
model designation
The model marked as the active, production model for a project. (The previous champion is kept as the
baseline so you can roll back to it.) See Model lifecycle for how these fit together.monitoring signal
A measure of how far a deployed model’s live behavior has moved from what it saw in training. AcuSight surfaces drift on dashboards and can raise alerts, signaling when it’s time to collect fresh data and retrain.
Devices & fleet
edge endpoint
A provisioned edge machine that captures images, runs a deployed model, and streams video. Devices are onboarded with a single setup command and a provisioning token, after which they appear on the Home page with live stats.
all your devices
The collection of devices in an organization, managed together. Models deploy to groups of devices, and monitoring spans the whole fleet.
set of devices
A named set of devices managed together as a deployment and operations target (shown under Edge Groups) — for example, “all cameras on line 3.” A related concept, Edge Stacks, deploys container stacks across a group.
per-device ops
Each device exposes its running containers, stacks, images, volumes, and networks for management directly in AcuSight, backed by Portainer — real container orchestration, not a façade.
onboarding credential
A single-use or multi-use token that authorizes a new device to join the platform. The setup script exchanges it for a device key.
Platform & integration
tenant boundary
The top-level tenant that owns everything — users, projects, devices, and data. AcuSight is multi-tenant: every resource is scoped to an org, and new users are routed to the right org automatically (usually by email domain).
AI assistant
AcuSight’s built-in AI assistant. It answers questions about your live data — read-only today — and is exposed over MCP so you can integrate it with your own agents.
vision-language model
A model that takes in an image and produces words about it — a description of what’s happening, rather than a label from a fixed list of trained classes. Slow and expensive to run compared to a detector, so it’s used selectively. See Vision Language Models for a demonstration; not part of a standard deployment.
image-and-text matcher
OpenAI’s model for measuring how well an image matches a given phrase. It’s an encoder, not a describer — it turns both the picture and the words into the same form and compares them, but writes no text of its own. Because you specify what to look for in plain words rather than training a detector for it, CLIP is quick to point at something new. In the demonstration it acted as a lookout, deciding which moments were worth handing to a VLM.
how AcuSight runs
How an AcuSight instance is run. The same product is available as a hosted service and as self-managed deployments — including on-prem and fully air-gapped installs for closed networks. Deployments differ mainly in sign-in and networking, not in features.
Want the deeper “why” behind these terms? The Concepts section explains how the pieces fit together. This page is the quick reference.