Description

LumiScan Annotation Web creates the training data for neural networks that detect and locate objects, for example for LumiScan AI Object Matching or LumiScan VGR. You draw an oriented bounding box (a rectangle that can be rotated) around each object in an image and give it a label, the class of the object. LumiScan Training learns from these boxes.

It is the successor of the desktop tool LumiScan Annotation. The annotation logic runs in a server; people annotate in their web browser, from any PC in the company network, without installing anything.

What it offers

Annotating

  • Oriented boxes drawn as center-line rectangles in three clicks, moved, resized and rotated with the mouse or the keyboard.

  • A magnifier with crosshair at the mouse pointer for edges to a tenth of a pixel.

  • Labels with keys 1 to 9, brightness and contrast, a region of interest, Save and copy to next for image series.

  • Undo and redo, and a history of every saved state of every image.

Working as a team

  • Projects: independent annotation jobs, each with its own images, settings and members.

  • Roles: annotator, reviewer, project admin and server admin. The server enforces them.

  • Sign-in with local accounts, the company directory (Active Directory / LDAP) or single sign-on (OpenID Connect, for example Keycloak or Microsoft Entra ID).

  • Edit locks: an image being edited is read-only for the others.

  • Revision notes: a reviewer leaves an instruction on an image for an annotator.

Quality

  • Several annotators per image: each annotates the image independently, without seeing the others' boxes.

  • The server compares their boxes. When all agree, it writes the reference annotation itself.

  • Reviewers see the differences numbered on the image and decide with one click.

  • Agreement statistics per annotator, per pair of annotators and per folder.

Speed

  • Box suggestions from your trained LumiScan networks (ONNX), or from an open-vocabulary detector for parts without a network yet. You accept, correct or discard them.

Data

  • The annotation files are byte-identical to those of the desktop tool: a CSV file with the boxes and a JSON file with the save history next to each image. LumiScan Training and other tools read them unchanged.

  • Projects in folders on the server or a network share, or in an S3 bucket (AWS S3, MinIO, SeaweedFS).

  • Export and import of whole projects, for example to send a project to HD Vision Systems for annotation.

  • Automatic backups of everything the server keeps.

How it works

The server (lumiscan-annotation-server, Windows or Linux) serves the web pages and the REST API. It reads the images, writes every annotation file, checks every input and enforces the roles. Its own data (accounts, projects, history, notes) is kept in a database file next to its configuration.

The browser shows the images and draws the boxes. Nothing is saved without the server.

The optional box suggestion service runs the networks on a computer with a GPU and answers the server’s requests. Images never leave the company network.

browser  --HTTP(S)-->  lumiscan-annotation-server  --->  project folders or S3 bucket
                              |                            <image>.png, .csv, .json
                              +--->  box suggestion service (optional, GPU)

Terms

Annotation

the boxes of one image, saved in <image>.csv.

Label

the class of an object, for example bracket. The labels of a project are shared by everyone in it.

Bottom edge

the edge of a box drawn in green. It gives the box its direction: an angle of 0 means the box top points up.

Reference annotation

with several annotators per image, the annotation training uses. Written automatically when the annotators agree, or decided by a reviewer.

Region of interest (ROI)

the part of the images that is annotated. Suggestions outside it are left out.