Data labeling for computer vision

Turn your footage into a trained model.
On your own PC.

SentinelLabeler is a Windows app for building object-detection datasets from video: pull frames, label them with AI help, catch the mistakes that ruin a model, and train it on your own graphics card. Your footage never leaves your PC.

Try free for 7 days → Explore all features
Runs on your PC • Nothing is uploaded • Cancel any time • Your projects stay yours if you stop paying
Interactive demo SentinelLabeler Canvas Preview
Click anywhere on the canvas to place a SAM 2 segmentation prompt.
Dataset Telemetry
TOTAL FRAMES
3,420
CLASSES
4 Classes
FORMAT
YOLO11 / ONNX
DATA PRIVACY
100% Local
In the app: P finds objects with AI, Q copies the previous frame's boxes, Space approves the frame and moves on.
The basics

What is a data labeler, and why does it matter?

A detection model learns from examples. Each example is an image plus boxes that say "this is a player", "this is a car". Those boxes are the labels, and a labeler is the tool you make them in.

1. FramesPictures from your video or camera
→
2. LabelsA box or outline around every object, with its class
→
3. TrainingThe model studies thousands of labeled frames
→
4. ModelFinds those objects in footage it has never seen

Labeling is where the hours go

Training a model takes minutes to hours of computer time. Labeling takes days of people time: thousands of frames, several boxes on each, and every box has to be right.

The model copies your mistakes

A missed object teaches the model to ignore it. A box with the wrong class teaches it the wrong name. Good labels matter more than a bigger model, so a labeler should help you find mistakes, not only make boxes.

How SentinelLabeler saves you time

Every slow part of labeling, and what we do about it

Video is the hardest thing to label by hand and the easiest to speed up, because one frame looks almost like the next. SentinelLabeler is built around that.

"The next frame is almost the same as this one."
  • Copy previous frame (Q): bring the last frame's boxes over and nudge them. New in 1.2
  • Fill between: label two frames, get estimated boxes on every frame between them. New in 1.2
  • Label similar: copy a frame's boxes onto near-identical frames you haven't labeled yet. New in 1.2
  • Track a box forward through up to 60 frames.
"Drawing every box by hand."
  • Find objects with AI (P): type your class names and let an open-vocabulary model draw the boxes, or use a model you trained yourself.
  • Click-to-mask with Segment Anything 2 for exact outlines.
  • Auto-label a whole batch of frames, then review.
  • AI boxes arrive as suggestions: nothing joins your dataset until you approve it.
"Mistakes I can't see until the model is bad."
  • Label Audit: a model checks your confirmed frames and lists likely mistakes, worst first: objects you missed, wrong classes, boxes with nothing in them. New in 1.2
  • One click to accept the fix, or Accept all on a frame. It never deletes your boxes for you, and every fix can be undone. New in 1.2
  • Quick checks with no model: tiny stray boxes, doubled boxes, boxes off the image. New in 1.2
"My score looks great but the model isn't."
  • Validation leak check: finds validation frames that are near-copies of training frames, which make a score look better than the model is. On one real 36,185-frame gameplay project it found 29% of the validation frames were near-copies. New in 1.2
  • Keep look-alike frames together when splitting, so the score means something. New in 1.2
  • A locked test set that every version is scored against the same way.
"I don't know what to label next."
  • Class check before training: which classes have too few boxes to learn from. New in 1.2
  • Results per class after training, and the class your model misses most, with how many boxes you have of it. New in 1.2
  • Start from an earlier model, from any project, instead of from scratch. New in 1.2
"Keeping a big project organized."
  • Label origins: see which boxes a person drew and which came from a model, a copy or an estimate. New in 1.2
  • Labeling guidelines per class, shown while you label. New in 1.2
  • Models page: every version, its scores, the classes it was trained on, and whether it beat the last one. New in 1.2
  • Project report: a one-page summary to send a client or teammate. New in 1.2

Items marked New in 1.2 arrived in version 1.2 (October 19, 2026). Current customers get it automatically through the in-app updater.

Built for real projects

Tens of thousands of frames, on one PC

Measured on a real 36,185-frame project with 35,560 confirmed boxes.

0.4 s
to scan 9,188 labeled frames for problems, using the saved label index
< 4 s
to write a full project report for 36,185 frames
YOLO · COCO · VOC
dataset export; trained models as ONNX
YOLO11 · YOLO26 · RT-DETR
boxes, masks, poses, rotated boxes and per-pixel segmentation
Private by design

Your footage never leaves your PC

Frames, labels and models are ordinary files in a folder you choose. Nothing you label is uploaded, and nothing is uploaded to train. The app goes online only to check your license (your key and a hashed ID of your PC), to check for updates, and to download the AI tools and models the first time you use them.

Your data stays yours

Stop paying, keep everything

If your subscription ends you can still open, browse and export every project you made. Labels are standard YOLO files and models are standard ONNX files, usable in any other tool.

What you need

Requirements

Windows 10 or 11, 64-bit.
An NVIDIA graphics card for training and the AI tools. Labeling by hand works on any PC; the AI tools also run on the processor, more slowly.
No setup: the AI tools install themselves on first run.
Disk space for your frames and trained models.
Updates

What's new

Version 1.2October 19, 2026
  • Label Audit: find likely labeling mistakes with a model, fix them in one click, undo any time.
  • Validation leak check, and a split that keeps look-alike frames together.
  • Fill between, Label similar and Copy previous frame for video.
  • Class check before training and results per class after it.
  • Label origins, labeling guidelines per class, and a shareable project report.
  • Models page: every model in one place, with what it was trained on.
  • Move frames, classes and models between projects; start training from an earlier model.
  • Rebindable class keys, including mouse side buttons and the wheel.
  • The AI tools pick the right PyTorch build for your graphics card and prove it works during install.
  • A "This PC" panel that shows what the app runs on, and a one-click PC report for support.
  • Reliability: fixes for class buttons, the Add class dialog, and labeling with your own model.
  • Security: stronger license checks, and the image decoder only opens JPEG, PNG, BMP, WebP and GIF files.
Version 1.1October 16, 2026
  • Silent in-app updates that keep all your projects and settings.
Simple Pricing

Unlimited labeling. Zero cloud limits.

Cancel any time. If your subscription ends you can still open and export every project, and every model you trained stays yours.

Monthly Plan
Full access, billed month to month. Cancel whenever.
$15/ month
  • Unlimited projects and frames
  • AI labeling: open-vocabulary detection, Segment Anything 2, your own models
  • Video tools: tracking, copy previous, fill between, label similar
  • Label Audit and validation leak check
  • Training on your own NVIDIA GPU: YOLO11, YOLO26, RT-DETR
  • Export to YOLO, COCO and VOC; models as ONNX
  • Every update while subscribed
Try Monthly free for 7 days →
2 Months Free
Annual Plan
Full access, billed once per year. Best value for creators.
$120/ year
  • Everything included in Monthly
  • Equivalent to just $10 / month
  • Priority updates & model architecture support
  • Priority support on Discord
Try Annual free for 7 days →
Secure Stripe checkout Cancel anytime with 1 click Runs 100% on your machine