Image and video annotation

Click the object.
The ink lays itself down.

One click segments an object, and one click follows it to the end of a clip. Around it sits the part teams actually live in: a pool of work, roles, and a queue that always knows whose turn it is.

Accounts are opened by an organization admin

One clip · 8 frames · 2 objects

Masked: 8 of 8

One clickIt follows the objectEvery frame saved

Human work: One click per object
Output: 8 annotation records

How a job moves

Five states, in the order they happen. Every screen below is a capture of the running product, not a drawing of one.

Upload
Pool
Assist
Track
Export

01Upload

Images, or a clip that becomes frames

Drop in images, or a video and a frame rate. Every frame becomes its own sheet, carrying the clip name, the frame index and its timestamp — so a clip is a queue, not a blob.

Takes: Images · video
Frame becomes: One sheet

Import · Video

Extraction runs through ffmpeg's fps filter, so frame i sits at i / fps seconds.

Frame rate

Clip: demo.mp4
Rate: 5 fps
Length: 8 s · example
Frames: 40
Each frame: One sheet

The marked frame, as a sheet

File: demo.mp4
Frame: 0018
At: 00:03.6
Rate: 5 fps

Illustrative extraction · demo.mp4

02Pool

Work handed out, not fought over

Sheets arrive in a pool. Hand out five or five hundred, to a person or to everyone, and watch them move: unassigned, annotating, completed. Organizations, projects, roles and permissions sit underneath.

Scopes: Org · project
States: Three

The Annotate section of a project, showing the task pool and the controls that hand work out — Project → Annotate · handing out a pool of 31 sheets

03Assist

Click the object; the ink lays itself down

One click segments the object. Right-click carves a piece back out. Give the shape a class and it prints in that class’s ink. Rectangle, polygon, circle, point, line and freehand are still there for when the model is wrong.

One click: ~0.1 s warm
Fallbacks: Six tools

The annotation screen with a mask traced around a cat, assigned to the class cat — Annotate · one click, one mask, inked as “cat”

04Track

Mark it once, follow it to the end of the clip

Point at the object on any frame and press Track. Labeloo carries it through the rest, masks landing frame by frame as they are computed. Several objects at once, each in its own ink — and the wait is shown as real frames and real seconds.

Per frame: 1.15 s on GPU
On a CPU: ~3 s

The video tracking screen mid-run, a bridge masked and the progress reading one of thirty frames — Video tracking · a run in progress, with its honest estimate

05Export

The records come back out

Every mark is an annotation record attached to its task, and the export endpoint hands them back. The dataset is yours; nothing about the format is ours to keep.

Output: Annotation records
Held in: Your database

Export format

A zip carrying the images themselves, plus one label file per image: the mask as a normalised polygon, not just a box.

yolo archive

zip

classes.txt
data.yaml
images/
- train · frame_0009.jpg
- val
- test
labels/
- train · frame_0009.txt
- val
- test

train / val / test split

train

val

test

Three numbers you set. The proportions drawn above are an example, not a default.

A yolo or coco export carries only the tasks that actually hold annotations.

Teams, roles, assignment

Who may do what, and who holds which sheet

A role is set twice — once on the organization, once on the project — and each scope carries its own flags. Work arrives as a pool, and every task records who it is assigned to and how far it has got.

Where a role lives

Organization

Organization role · 8 flags

Project

Project role · 5 flags

Task

Carries: Assignee, status

Annotation

What a seeded role may do

Organization roleSet on the organization
Flag	Admin	User
Admin	True	False
Edit organization	True	False
Delete organization	True	False
Edit members	True	False
Edit roles	True	False
Edit projects	True	True
Create projects	True	True
Delete projects	True	False

Project roleSet on the project
Flag	Admin	User
Edit project	True	False
Delete project	True	False
Edit members	True	False
Edit roles	True	False
Upload files	True	False

TrueEmpty cell, false

How work is handed out

UnassignedAnnotatingCompleted

Assigned to: Member A
Assigned to: Member B
Assigned to: Member C

Handed out: In bulk, or claimed
Bulk asks: How many sheets, and who
Task carries: Assignee, status

What the model costs you in time

Tracking is not instant and the interface never pretends it is: a run shows the frames done and the seconds left. These are measurements from one machine, not a promise about yours — the model runs on a service you point the organization at, so the number moves with the hardware behind it.

Single click, image: ~0.1 s; once the sheet is embedded
Video, per frame, GPU: 1.15 s; Apple GPU, sam2.1 small
Video, per frame, CPU: ~3 s; the same run without a GPU

Put it on your own data

Labeloo runs on your own machines, with the segmentation service addressed per organization — so the model can sit on a laptop today and a GPU box tomorrow without touching the app. Tell us what you are labelling and we will get you set up.

Try it with a demo account

Email: demo@labeloo.app
Password: LabelooDemo2026
Signs in as: Member, with six sheets waiting

Click the object.The ink lays itself down.