Label Audit
A model re-checks your confirmed frames and lists likely mistakes, worst first: missed objects, wrong classes, empty boxes. Accept a fix in one click; every fix can be undone.
Desktop app for Windows · Version 1.2
From raw video to a trained detection model. On your own PC.
Labeling is the slow part of computer vision. SentinelLabeler drafts the boxes with AI, puts the uncertain ones in front of you first, and trains the model on your own graphics card. Your footage never leaves your computer.
7 days free, then $15/month or $120/year. Card required; cancel before day 7 and you pay nothing.
270 boxesdrafted on 16 street photos in about 25 seconds, on a laptop RTX 5050
0.4 sto scan 9,188 labeled frames for problems in a real 36,185-frame project
0 uploadsframes, labels and models stay in a folder you choose
YOLO · COCO · VOCstandard exports; trained models as ONNX
Real output, not a mockup
See exactly what the AI hands you to review.
These are the boxes SentinelLabeler's batch auto-label drew on three public photos, exported straight from the app. Hide a class or raise the confidence threshold to see what you would review first.
Photo
Classes
Minimum confidence
Showing boxes at 0.20 or higher
On screen
0
Model: YOLO-World (yolov8s-worldv2) through SentinelLabeler 1.2 batch auto-label, prompted with five class names. Boxes are unedited.
How it works
Four steps, one app.
No scripts between tools, no uploading footage to a website, no converting formats by hand.
01
Bring frames
Drop in videos or image folders. Frames are pulled out, near-duplicates skipped, and blurry ones can be left out.
02
AI drafts labels
Type your class names and an open-vocabulary model draws the boxes, or use a model you trained.
03
You review fast
Worst-confidence frames first. Accept or reject a whole page of suggestions at once.
04
Train and export
Train YOLO or RT-DETR on your GPU, then export the dataset or an ONNX model.
Annotate
AI draws the first draft. You make the call.
Press P and an open-vocabulary detector finds every object you named. Press W and Segment Anything 2 turns one click into an exact outline. Suggestions are dashed until you approve them, so nothing the AI guessed reaches your dataset by accident.
- Find objects with AI from plain class names, or with your own trained model
- Click-to-mask with Segment Anything 2.1
- Boxes, polygons, poses, rotated boxes and per-pixel masks
- Keyboard first: one key per class, Space approves and moves on
Bulk Review
Review a page of frames at a time.
Bulk Review shows every frame the AI touched, sorted with the least confident first, because that is where its mistakes are. Click a box to accept or reject it, then apply the whole page.
- Worst confidence first, or filter below any threshold
- Accept all or reject all on a page, then fix the exceptions
- Turbo Review: flip through frames with the arrow keys for a fast visual check
Train Studio
Train on your graphics card, with a sanity check first.
Train Studio finds your GPU, measures your box sizes, and tells you which classes are too thin to learn before you spend an hour on a run that was never going to work.
- YOLO11, YOLO26 and RT-DETR, from scratch or from an earlier model
- Class check before training, results per class after it
- A locked test set so every version is scored the same way
- Export to ONNX, ready for your own app
Dataset quality
Finds the mistakes that quietly ruin a model.
A missed object teaches a model to ignore it. A near-duplicate in your validation set makes a weak model look strong. SentinelLabeler checks for both.
Validation leak check
Finds validation frames that are near-copies of training frames. On one real 36,185-frame project, 29% of the validation frames were.
Class check
Before training: which classes have too few boxes. After training: results per class, and the one your model misses most.
Label origins
See which boxes a person drew and which came from a model, a copy or an estimate between two frames.
Video tools
Copy the previous frame, fill boxes between two labeled frames, label look-alike frames, and track a box forward up to 60 frames.
Project report
A one-page summary of classes, boxes, splits and model scores to send to a client or teammate.
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, check for updates, and download the AI models the first time you use them.
No lock-in
Stop paying, keep everything.
If your subscription ends you can still open, browse and export every project. Labels are standard YOLO, COCO or VOC files and models are standard ONNX files that work in any other tool.
Specifications
What it runs on, what it reads and writes.
| System | Windows 10 or 11, 64-bit |
|---|---|
| Graphics card | NVIDIA GPU for training and fast AI labelingLabeling by hand works on any PC; the AI tools also run on the processor, more slowly. |
| Label types | Bounding boxes, polygons, pose keypoints, rotated boxes (OBB), semantic masks |
| AI labeling | YOLO-World open-vocabulary detection, Segment Anything 2.1, your own trained models |
| Training | YOLO11, YOLO26, RT-DETR on your own GPU |
| Export | YOLO, COCO, Pascal VOC datasets; models as ONNX |
| Input | Video files and image folders (JPEG, PNG, BMP, WebP, GIF) |
| Setup | One installer. The AI tools install themselves on first run, matched to your graphics card. |
Pricing
One plan, every feature.
Unlimited projects and frames. Start with 7 days free.
Monthly
$15/ month
Billed monthly. Cancel any time.
- Every feature, unlimited projects
- AI labeling and training on your GPU
- Every update while subscribed
Yearly2 months free
$120/ year
$10 a month, billed once a year.
- Everything in Monthly
- Priority support on Discord
- Every update while subscribed
Card required for the trial. Cancel before day 7 from Account → Manage billing and you are not charged. Full pricing details
Questions
Before you try it.
Does anything get uploaded?
No. Your frames, labels and models stay in a folder on your PC. The app only contacts our server to check your license (your key and a hashed ID of your PC) and to check for updates.
Do I need an NVIDIA graphics card?
For training, yes. Labeling by hand works on any Windows PC, and the AI labeling tools also run on the processor, just more slowly.
Can I use my labels in other tools?
Yes. Export to YOLO, COCO or Pascal VOC, and trained models export as ONNX. Your projects stay readable and exportable even if your subscription ends.
How does the free trial work?
New customers get 7 days free on either plan. You add a card at checkout and nothing is charged until day 7. Cancel before then from Account → Manage billing and you pay nothing. One trial per customer.
Is it only for gameplay footage?
No. It started with gameplay and camera footage, which is why the video tools are strong, but it works with any images: street scenes, products, wildlife, inspection photos.
Label your next dataset in an afternoon, not a week.
7 days free. Runs on your PC. Cancel any time.