Turn video into training data.

Bounding boxes, polygons and keypoints in the same project, on the same video.

• Boxes for object detection, polygons for instance segmentation, keypoints for pose estimation
• Each exports to the matching YOLO format
• Every object carries its own label and colour

Geometry is stored in native video pixel coordinates, so it is exact regardless of how you zoom.

• Drag to move, eight handles to resize, drag individual polygon vertices
• Double-click an edge to insert a vertex, right-click one to remove it
• Mouse-wheel zoom, middle-drag pan, and a Fit control
• Small clips scale up automatically, so you never annotate a postage stamp

Built for hours of labelling, not for demos. No pre-extracting frames to a folder — import an MP4 and start.

• Space to play or pause, arrow keys to step, Shift for ten frames
• K to pin a keyframe, Delete to remove one, Esc to cancel
• Scrub, play and step frame by frame while you work

Mark an object where it appears, move to a later frame and adjust it — every frame between is computed automatically and smoothly.

• Label a 300-frame sequence with four keyframes instead of 300 boxes
• Interpolation is linear and entirely under your control — no model in the loop
• Add or remove a polygon vertex and the change propagates across every keyframe of that track

Each object is a track with a stable identity and label, not a pile of unrelated per-frame rectangles.

• The object panel says what you are looking at on this frame: keyframe, interpolated, or held
• Hold-forward tracking keeps an object on screen after its last keyframe, so you can refine it later
• End the track when the object leaves the scene — no ghost labels polluting the dataset

One lane per object with diamond markers at every keyframe.

• Click a diamond to jump to it, right-click to delete it
• The span bar shows exactly which frames each object exists on
• Dashed outlines on the frame mean interpolated — computed, not hand-drawn

One click produces the standard YOLO layout that common training pipelines accept directly.

• images/train, images/val, labels/train, labels/val and data.yaml
• Point your trainer at the data.yaml and start training
• YOLO is the only export format in v1.0 — no COCO or Pascal VOC

Detect, segment or pose — chosen at export time, from the same annotations.

• Detect — class cx cy w h; every shape becomes a bounding box
• Segment — polygon outlines; boxes are emitted as rectangles
• Pose — with kpt_shape and flip_idx written into data.yaml

Export writes real JPG frames next to their label files, with matching filenames. The source video is never needed downstream.

• The dataset folder is self-contained — zip it and send it anywhere
• Choose which videos to include, set the train/validation split, apply a frame stride
• Progress reported per frame, and the export can be cancelled