EI Label Studio

A Vercel-hosted Label Studio wrapper for Edge Impulse. Connect an Edge Impulse project, pull its data samples, relabel them in an embedded Label Studio canvas, and push every correction straight back to Edge Impulse — all driven by shareable URLs.

🔗 Live: https://label.jennyspeelman.dev

Edge Impulse Label Studio with Segment Anything annotation

Features

• Connect with an API key — paste a project API key (ei_…) + project ID. The key is kept in a secure, http-only cookie and proxied server-side; it never touches the URL or local storage.
• Image, audio & time-series — labeling templates that map onto Edge Impulse modalities: image classification, bounding boxes, audio classification, and multi-axis time-series labels. The modality is auto-detected per sample, or you can force a template.
• Round-trips to EI — samples are pulled via the Studio API; the current label is shown as a pre-annotation; submitting relabels the sample back in Edge Impulse.
• Interactive segmentation (SAM) — a Segment Anything template enables Point2Label / Bbox2Label: click or drag and a hosted SAM 2.1 backend on Beam Cloud returns a mask + bounding box. The mask stays in the canvas; the bounding box is pushed back to Edge Impulse. See ml-backend/.
• URL-driven — deep-link presets (project, category, template, filters, theme, embed…) in the spirit of synthetic-data-studio. See docs/url-parameters.md or the in-app /docs page.
• Embeddable — ?embed=1 strips the chrome for iframe use.

How it works

Browser --POST /api/ei/session--> http-only cookie (apiKey + projectId)
   |
   |- GET  /api/ei/projects        -> Studio  GET /v1/api/projects
   |- GET  /api/ei/samples         -> Studio  GET /v1/api/{projectId}/raw-data
   |- GET  /api/ei/media/{p}/{s}   -> Studio  /raw-data/{s}/{image|wav}  (or synth CSV for series)
   |- POST /api/ei/relabel         -> Studio  POST /raw-data/{s}/rename
   \- POST /api/ei/upload          -> Ingestion POST /api/{category}/data

All Edge Impulse traffic is proxied through same-origin Next.js Route Handlers (Edge Impulse blocks browser CORS). The Label Studio Frontend runs entirely client-side — there is no Label Studio server. Its bundle is vendored under public/vendor/label-studio.

Interactive segmentation (SAM)

The Image · Segment Anything (SAM) template adds SAM-style Point2Label / Bbox2Label. The PyTorch + SAM stack can't run on Vercel (no containers; functions cap at 500 MB), so the model is hosted separately and called server-side:

Canvas (smart tool) --> POST /api/ei/predict --> SAM backend (Beam Cloud GPU)
                              |  fetch image from EI, hand backend a temporary Blob URL
                              \- returns mask + bbox; the box is saved to EI on submit

Setup:

1. Deploy the backend — a GPU-enabled Beam Cloud deployment; see ml-backend/.
2. Link a Vercel Blob store to the project (Storage tab) — this sets BLOB_READ_WRITE_TOKEN, used to give the backend a temporary public URL for each image.
3. Set SAM_BACKEND_URL to the Beam deployment's /predict URL and SAM_BACKEND_AUTH to the Bearer Token. See .env.example.

Local development

pnpm install
pnpm dev
# open http://localhost:3000

Get a project API key from your Edge Impulse project Dashboard -> Keys, and the numeric project ID from Project info.

Tech

Next.js (App Router) · TypeScript · Tailwind CSS v4 · shadcn/ui · Zustand · Label Studio Frontend. Deployed on Vercel.

Notes & limitations

• An Edge Impulse API key is scoped to a single project, so the project picker shows just that project. (Account-wide listing would require JWT login.)
• For detect (bounding-box) templates, edited boxes are pushed back to Edge Impulse via the Studio API. For time-series templates, the sample-level label is updated, but per-segment annotations are not yet pushed back.
• For SAM, the mask is an in-canvas labeling aid only — Edge Impulse stores bounding boxes, so the mask's bounding box is what gets saved. Requires the hosted backend (ml-backend/).
• The standalone Label Studio Frontend npm package no longer ships a built bundle, so a known-good build is vendored in this repo and loaded on demand.