Bounding boxes
Classic object detection, with confidence shown on every suggestion.
Features
Everything between your footage and a working model.
One desktop app for collecting frames, labeling them with AI help, checking the dataset, and training and comparing models. All of it runs on your PC.
01 · Label
Fast labeling, with AI doing the first pass.
Every tool has a key, and AI suggestions stay dashed until you approve them.
- Find objects with AI P: an open-vocabulary detector boxes every class you named
- Click-to-mask W: Segment Anything 2.1 turns a click into an outline
- Edge snap S: tighten a rough box to the object's edges
- Track T: follow an object into the next frames with optical flow
- Rebindable keys, including mouse side buttons and the wheel
Polygons
Instance segmentation outlines, drawn or made from one SAM click.
Pose and rotated boxes
Keypoints for pose models, oriented boxes for aerial and top-down footage.
Semantic masks
Per-pixel labels for scene segmentation.
02 · Video
Built for video, where frames repeat.
One frame of video looks almost like the next. These tools use that, so you label an object once instead of fifty times.
Frame extraction
Pull frames from video at the rate you choose, skip near-duplicates, and leave out blurry frames.
Copy previous Q
Bring the last frame's boxes over, then nudge the ones that moved.
Fill between
Label two frames; get estimated boxes on every frame in between.
Label similar
Copy a frame's boxes onto near-identical frames you haven't labeled yet.
Track forward
Follow a box through up to 60 frames.
Live Capture
Record training footage straight from a capture card or screen into a project.
03 · Review
Check a thousand suggestions without opening a thousand frames.
- Bulk Review: a page of frames at once, least confident first; accept or reject per box, then apply
- Turbo Review: flip through frames with the arrow keys for a fast visual check
- Auto-approve high-confidence suggestions in a batch run, so you only review the doubtful ones
04 · Dataset quality
Catch the problems a training score hides.
Label Audit
A model re-checks confirmed frames and lists likely mistakes, worst first. One click accepts a fix; every fix can be undone.
Quick checks
Without a model: tiny stray boxes, doubled boxes, boxes that run off the image.
Validation leak check
Finds validation frames that are near-copies of training frames, and a split that keeps look-alike frames together.
Class check
Shows which classes have too few boxes to learn, before you train.
Label origins
Which boxes a person drew, and which came from a model, a copy or an estimate.
Guidelines per class
Write what counts as each class once; it shows while anyone labels.
05 · Train and compare
Train on your GPU. Know if the new model is actually better.
- YOLO11, YOLO26, RT-DETR, from scratch or from an earlier model in any project
- Results per class, and the class your model misses most
- A locked test set, so every version is scored the same way
- SAHI tiled inference for small, far-away objects
- Model Arena: overlay two ONNX models on the same frames
- Burst Frames: run up to four models over a batch at once
- Models page: every version, its scores, its classes, and whether it beat the last one
06 · Projects and files
Standard files, organized projects.
Export
YOLO, COCO and Pascal VOC datasets; trained models as .pt and ONNX.
Import
Bring in existing datasets, including Roboflow exports, and adopt their classes.
Move between projects
Copy frames, classes and models from one project to another.
Project report
A one-page summary of classes, boxes, splits and model scores to share.
Local files
Every project is a normal folder on your disk. Back it up or move it like any other folder.
Self-updating
Updates install from inside the app and keep every project and setting.