Release Notes

Sizes and percentages are written the way the chosen language writes numbers, for example 2,0 GB and 96,6 % in German.

German manual: all screenshots and videos show the German interface, and the names of buttons and menus are the ones the German interface shows.

Some texts still appeared in English in every language: the reasons for differences in review mode (for example "overlap only 74 %"), the Delete buttons, the number of open revision notes, the hint in Members & access, the width and height of the region of interest, the text and status of Rate with network, and the error messages for uploads and images.

Interface in five languages: English, German, Spanish, French and Chinese (simplified). The app starts in the language of the browser, English if there is none for it.

Language saved with your account: choose a language on the login page or under your name → Language. Your choice applies in every browser and on every PC you sign in with. Same as the browser goes back to the browser’s language.

Own texts and languages without a new version: administrators can correct texts, use the terms of their company or add a language by putting a translation table into the server’s translations folder. The change applies at the next reload of the page, without a restart. See Languages and Translations.

Dates and times are shown the way the chosen language writes them.

Most messages of the server (for example "The current password is wrong.") appear in the chosen language too.

Before an update changes the database format, the server now keeps a copy of the old database next to it (for example annotation-server.db.v3.bak). Going back to the previous version then only means installing it and putting the copy in place of the database.

New configuration setting translations_dir (see Configuration).

An image that was read-only because someone else, or another window of your own, had it open stayed read-only until it was opened again. It now becomes editable by itself as soon as it is free, and the app says so.

A window that had lost its edit lock to someone else could take the lock back in the background while it still showed the image read-only. The image was then blocked for everyone. Fixed.

Downloads with umlauts, accents or Chinese characters in their name (exports, project files, backups) could arrive with a garbled file name in some browsers. The exact name is now always kept.

Faster start: the app is loaded while the translations are fetched, and the language chosen before is fetched together with the list of languages.

Unchanged translation tables are not sent again: the browser uses its copy.

Adding images from the browser needs one listing of the folder less per file, which saves many requests in large S3 buckets.

The database format changes (version 4). Older versions cannot open the updated database. The server keeps the old one as annotation-server.db.v3.bak next to it: to go back to 2.3, stop the server, install 2.3 and replace annotation-server.db with that copy (changes made with 2.4 in the meantime are then lost).

JPEG, BMP and TIFF besides PNG, including multi-page TIFF: every page is an image of its own with its own annotation file (see Image Formats). 16-bit images are shown with the brightness range they actually use (10, 12, 14 or 16 bit).

Adding images in the browser (project admins): + Add images above the image list takes files, a whole folder or drag and drop, into an existing or a new folder. Large files are sent in pieces. Works for folders on the server and for S3 buckets (see Adding Images).

Mark as done: Mark done next to Save, or Shift+D to save, mark and go on, marks an image as finished, also when it has no boxes. The image list shows a ✓ and the count, and Show filters by it.

COCO export: Export offers the COCO format, used by CVAT, Label Studio, FiftyOne and many training frameworks: images with annotations as ZIP, or only the annotations as JSON for one set or all (see Exporting the Annotations).

Set filter: the image list shows in which set (training, validation, test) the export puts each image and can filter by it.

Keep each folder together is remembered with the export settings of the project.

The built-in networks for box suggestions read all image formats.

Done marks are included in the project export and import.

New configuration setting upload_max_mb for the largest file that can be added in the browser.

TIFF pages are converted for the browser once and then kept in a cache; the page list of a TIFF file is only read again when the file changes.

Added images appear in the list at once, without waiting for the next scan of the folder.

The database format changes (version 3). After the update, older versions can no longer open the database.

Single-PC installation: a Windows installer with a start-menu and desktop shortcut. LumiScan Annotation runs in its own window, without a login, and only accepts connections from the same PC (see Single-PC Installation).

Copying boxes: take the boxes of the previous image (V), copy boxes to the next images of a series, and copy and paste single boxes or all boxes with Ctrl+C / Ctrl+V, also between images (see Copying Boxes and Templates).

Box templates: save a box as the typical size and orientation of its label; T places one at the mouse pointer.

Finding images: search by file name, filter by state, by label or by the uncertainty of the network, and sort the list (see Finding and Sorting Images).

Active learning: Rate with network lets the project’s network rate the images; Network least sure first shows the images where annotation helps most.

Export for LumiScan Training: a ready LumiScan Training project as ZIP, with a stable split into training, validation and test set.

Built-in box suggestions: trained LumiScan networks run inside the server, without a separate service and without Python.

Update check: project admins see when a newer version is available (single-PC installation; optional on servers).

Closing the window asks about unsaved changes on macOS and Linux too (before: Windows only).

Windows: configuration and data folders below user or folder names with umlauts did not work.

The image list knows the box count and the labels of every image from the image index, so filtering by label does not read the annotation files again.

Box suggestions from the built-in networks need no round trip to a separate service.

Sign-in and roles: local accounts, Active Directory / LDAP and single sign-on with OpenID Connect (Keycloak, Microsoft Entra ID, Okta and others); roles annotator, reviewer and admin, checked by the server (see Users and Roles, Sign-in Methods).

Projects: independent annotation jobs with their own folder, settings, members, history and notes; export and import of a whole project as one file (see Projects).

Projects in S3 buckets: AWS S3, MinIO, SeaweedFS and other S3-compatible stores (see Projects in S3 Buckets).

Several annotators per image: blind annotation, automatic reference when all agree, a review mode with the differences numbered, and agreement statistics (see Several Annotators per Image).

Version history and undo: every saved version of an annotation can be viewed over the image and restored; undo and redo while editing (see History and Undo).

Revision notes: instructions on an image for a person, pinned to a box if needed, with replies (see Revision Notes).

Box suggestions: a service with Grounding DINO and SAM 2, or with trained LumiScan networks, suggests boxes (P) (see Box Suggestions).

Precise drawing: magnifier with crosshair and pixel grid, center lines through every box, a second drawing mode (edge rectangle, E), and turning the bottom edge of a box (R).

Backup and restore of the database and the configuration, automatic and on demand (see Backup and Restore).

HTTPS by default with a self-signed certificate that the server creates and renews itself.

Responses are compressed for browsers that accept it.

New login page with the LumiScan apps.

A / D skip images hidden by the filter of the image list.

A second server on the same port shared the port silently on Linux; the server now refuses to start and says why.

The read-only banner said "reviewing" in projects with one annotator per image, where the other person is annotating; it now says "editing".

Drawing, moving boxes, panning and the magnifier are smooth also with many boxes: about five to nine times less work per mouse movement.

Image lists and statistics come from an index in the database instead of reading every file again.

Images from S3 buckets are cached on the server and only downloaded again when they change.

A server that listened on all addresses with plain HTTP now answers with HTTPS on the same port. Set "https": "off" in the configuration to keep HTTP.

The access code (auth_token) of version 2.0 is replaced by sign-in; a configuration that still contains it is refused with a hint.

Annotation in the browser, served by a server for Windows and Linux.

Annotation files byte for byte as written by the desktop tool, so LumiScan Training reads them unchanged.

The same way of drawing oriented boxes as in the desktop tool, with its keyboard shortcuts, brightness and contrast, region of interest and labels on keys 1 to 9.

Several people at once: edit locks per image, a read-only view for images others have open, and a warning when a file was changed by someone else.