What we're building

Object Detection, Instance Segmentation, Pose Estimation, Face Detection, Classification — each with dedicated toolset and export filters.

YOLO v5–v12, COCO, Pascal VOC, LabelMe, CreateML, CVAT, Darknet, Florence-2, Multiclass, CLIP, PaliGemma, RetinaNet and more.

Batch detection and polygon mask generation with any YOLOv8/YOLO11 ONNX model. Class names read from model metadata.

Batch skeleton prediction and 5-landmark face detection with YOLOv8-Pose / YOLOv8-Face ONNX models.

21-point hand skeleton with per-finger coloring and named joints. Body / Hand model switch in the Pose dialog. Dynamic keypoint counts — any YOLO-pose model works.

New YOLO Pose handler with full round-trip import. COCO keypoint export/import fixed — skeleton edges and keypoint names included in categories.

Settings → Models with one path per task (Detection, Segmentation, Body, Hand, Face). Relative defaults survive reinstalls; Full installer bundles all models.

Theme selector in Settings → General, persisted across sessions. Dark theme contrast fixed in dialogs and tab bars.

One-time license with machine activation/deactivation, offline grace period, and a full-featured 14-day trial.

Frame-by-frame annotation of MP4/AVI files with scrubber, jump controls and slideshow playback.

Live HUD showing labeled progress, annotation count, and top-5 class distribution directly on the canvas.

58 QTest unit + integration tests covering all format round-trips, dataset CRUD, and project save/load.

When a project is loaded and image files are missing from disk, show a "Verify Paths" warning instead of silently showing empty entries.

Save and restore window size and position across sessions using QSettings saveGeometry / restoreGeometry.

Check version.json on startup and show a status bar notification with a download link when a newer version is available.

Autosave interval, default export format, and UI language options in Settings → General.

Let users adjust segmentation mask opacity from Settings so the underlying image remains as visible as they need.

Attach arbitrary key-value metadata to individual annotations (e.g. "occluded: true", "truncated: false", "difficulty: hard").

Multi-select images in the file list to delete, move to a different split, copy annotations, or export a subset.

Filter the file list by class name, annotation count, labeled/unlabeled status, or review flag.

Store per-keypoint confidence values from AI inference. Export in COCO v2 format with visibility flags 0/1/2.

Flag class imbalance, missing annotations on images, or unusually small/large bounding boxes in the statistics panel.

Qt6 cross-platform build targeting macOS (arm64 + x86_64) and Ubuntu LTS. Same feature set as Windows.

Trigger a training job (local GPU or cloud) directly from the export dialog and import the resulting model back for auto-labeling.

Share a project with reviewers who can approve, reject or comment on individual annotations without editing them.