Most video annotation tools ask for two things before you label a single frame: an account and an upload.
For a quick experiment, a sensitive clip, or a one-off dataset, both are friction.
LabelOp Tracker is a free video annotation and multi-object tracking app that runs entirely in your browser tab. You open a video, let a detector and tracker propose boxes, fix what is wrong, and export a dataset. Your video and your annotations stay on your computer.
This guide covers what Tracker is, who it is for, a full walkthrough from opening a video to exporting, and the limits you should know before you start.
What LabelOp Tracker is
Tracker is part of the LabelOp family, next to the LabelOp labeling platform, LabelOp Moments and LabelOp Retouch.
Its job is narrow: turn a video clip into tracked, reviewed annotations you can train on.
What it does:
- Auto-annotate with a detector plus a multi-object tracker (ByteTrack), running on your GPU through WebGPU
- Manual drawing of boxes, polygons and ellipses, with keyframes between frames
- Track editing: split, merge, fill gaps, remap classes, delete ranges
- A QA tab that flags likely problems such as ID changes and duplicate boxes
- Exports to COCO JSON, YOLO TXT, YOLO-SEG, YOLO-POSE and YOLO-CLS, packaged as a ZIP
What it does not need:
- No account
- No server-side processing of your video
- No API keys
Projects and annotations are stored in this browser (IndexedDB, with video files in the Origin Private File System). Videos and annotations are never uploaded. The site does use privacy-friendly page analytics, and the AI model weights are downloaded once on first use.
Who it is for
Tracker fits well when:
- You are an ML engineer or researcher who needs a few hundred tracked frames to test an idea
- You work with footage you cannot send to a third-party server
- You want to check how a COCO-class detector performs on your own clips before committing to a labeling project
- You need COCO or YOLO files quickly and do not want to set up a self-hosted tool
It is a single-user, single-browser tool. If you need several annotators, reviewer assignments or dataset versions, that is what LabelOp is for (more on that at the end).
Requirements
Check these before you start:
| Requirement | Detail |
|---|---|
| Browser | Chrome or Edge recommended |
| Browser APIs | WebCodecs for video decoding, WebGPU for auto-annotation |
| Video formats | MP4, M4V or MOV. HEVC converts automatically |
| First AI run | One-time model download (RF-DETR nano is about 103 MB), cached afterwards |
When you open the app, a status line tells you whether your browser is ready. If WebGPU is missing, you can still open video and draw by hand, but the detector has no CPU fallback, so auto-annotation will not run.
Walkthrough: from video to dataset
Step 1: Open a video
Click Open video, or drop a file onto the page.
There is no separate "create project" step. Choosing a video creates the project immediately, named after the file. The project then appears in your local project library, so you can close the tab and come back later.
Step 2: Run Auto-annotate
A new project guides you to run Auto-annotate first. It lives in the Setup tab.
The defaults are sensible:
- Model: RF-DETR nano, which detects the 80 COCO classes (person, car, bicycle, dog and so on). The picker shows the download size before anything downloads.
- Start auto-annotate: runs detection on each frame and links detections into tracks.
Under More options you can:
- Test on this frame to check settings on a single frame before a full run
- Set a Minimum confidence (default 0.35; lower finds more objects but adds more false detections)
- Limit detection to specific classes
- Choose a frame range
One run covers up to 2,000 frames. On a longer video, Tracker asks you to pick the range.
Progress shows the model download first, then frame-by-frame tracking. You can cancel at any point.
For a deeper look at how detection and tracking interact, read how auto-annotation works in Tracker.
Step 3: Look at the result
After the run, the Tracks tab lists every track with its class and duration. You can sort by first appearance, duration or track ID, hide a track, or "solo" one to see it alone.
The timeline under the video shows one lane per track, plus an overview chart of how many objects were detected and how confident the detector was across the clip. Click the chart to jump to a frame.
Useful view toggles in Setup: show or hide boxes, names and motion trails, and color boxes by class instead of by track.
Step 4: Fix the tracks
Auto-annotation is a first draft. Typical fixes:
- Wrong class: select the track and change its class. The change applies to every frame of that track.
- One object, two IDs: merge one track into the other from the Tracks tab, or fill the gap between two runs of the same track from the timeline.
- Two objects, one ID: put the playhead where the switch happens and press
Sto split the track. - Bad box on one frame: drag it, nudge it with the arrow keys, or press
Deleteto remove the box on that frame only. - Junk in part of the clip: Shift-drag on the timeline to select a range, then delete everything inside it or keep only that range.
To add an object the detector missed, choose a draw tool (2 box, 3 polygon, 4 ellipse) and draw it. Add keyframes on later frames and Tracker fills the frames in between. For hand-drawn objects, the Follow video motion option makes boxes follow the motion in the video between keyframes instead of sliding in a straight line.
Every edit can be undone with Ctrl+Z (⌘+Z on Mac).
Step 5: Review in the QA tab
The QA tab builds a review queue from four signals:
| Signal | What it suggests |
|---|---|
| ID change | A track ended and a new one started nearby. One object may have a new ID. |
| Duplicate | Two tracks cover almost the same area. The same object may be labeled twice. |
| Sudden jump | A box moved or resized unusually between frames. |
| Unsure | The model was not confident about this object. |
Step through the queue with J and K, mark items Fixed (F) or Not an issue (X), and filter by problem type. If nothing is flagged, it is still worth skimming the video before you export.
If you have MOT-format ground truth, the QA tab can also compare your tracks against it and report MOTA, HOTA, IDF1 and ID switches.
Step 6: Export
The Export tab starts with a short checklist: open review items, object and class counts, and hidden objects (which are still exported).
Then pick a format:
| Format | What you get |
|---|---|
| COCO JSON | annotations.json with boxes, polygons, keypoints and a track_id on each annotation |
| YOLO TXT | One label file per frame plus classes.txt |
| YOLO-SEG | Polygon labels from polygons, ellipses and segmentation masks |
| YOLO-POSE | COCO-17 skeletons, train/val lists and data.yaml |
| YOLO-CLS | Whole-frame labels in train/val class folders |
You can include decoded JPEG frames, choose a frame range, and on supported browsers save large ZIPs directly to disk. The export guide for tracked video covers each format in detail.
Keyboard shortcuts worth learning first
| Key | Action |
|---|---|
Space |
Play / pause |
← / → |
Previous / next frame (with no track selected) |
Shift + ← / → |
Jump 10 frames |
1 to 7 |
Switch tool: select, box, polygon, ellipse, segment assist, select similar, pose |
S |
Split the selected track at the playhead |
Delete |
Delete the selected box on this frame |
J / K, F, X |
Move through the review queue and mark items |
Canvas shortcuts need the video focused. Click the video once after using the side panel.
Limits to know
Be clear about what Tracker is and is not:
- One browser, one person. Projects live in this browser's storage. There is no sharing or multi-user review.
- Clearing site data removes projects. Tracker asks the browser for persistent storage, but the browser may decline. Save a project backup (
.labelop.zip) for anything you care about. Backups must be smaller than 4 GB; for larger videos, save an annotations-only backup. - 2,000 frames per auto-annotate run. Longer videos need multiple ranges.
- The default detector knows COCO classes. For other objects, use the OWLv2 text-prompt model (about 295 MB) or upload your own ONNX detector.
- Pose is manual. You can place COCO-17 skeletons, but no pose-estimation model ships yet.
- Train/val splits from one video are not independent. Tracker offers a frame gap before validation, but distinct source videos are the right way to build an evaluation set.
When to move to LabelOp
Tracker gets one person from a clip to a dataset. Teams usually need more:
- Several annotators and reviewers with roles (team collaboration)
- Dataset history and version comparisons (dataset management)
- Exports to more formats from a shared project (export formats)
A common path is to prototype with Tracker, export JPEG frames with COCO or YOLO labels, and bring that dataset into a LabelOp project once review and versioning matter.
If you are planning a larger video project, the video annotation workflow guide covers sampling and ID rules that apply whatever tool you use.
Try it
Open LabelOp Tracker in Chrome or Edge, drop in a short clip, and run Auto-annotate. A 30-second traffic or pedestrian clip is a good first test: it shows tracking, ID changes and the QA queue within a few minutes.
FAQ
Is LabelOp Tracker really free?
Yes. It is free to use and needs no account. Processing runs on your own computer, in your browser, so there is no server-side compute to pay for.
Does Tracker upload my video?
No. Videos and annotations are never uploaded. They are stored in your browser's local storage. Model weights are downloaded once and cached, and the site uses privacy-friendly page analytics.
Which browsers does Tracker support?
Chrome or Edge are recommended. Tracker needs WebCodecs to decode video and WebGPU to run the auto-annotation detector.
What video formats can I open?
MP4, M4V and MOV. HEVC video converts automatically when you open it.
Can I use Tracker with a team?
Tracker is single-user and stores projects in one browser. For shared projects, review and dataset versioning, continue in LabelOp.