labelimgplusplus 4.0.0rc0

Core Annotation Features

• Five annotation formats: PASCAL VOC, YOLO bbox, CreateML, COCO, and YOLO-seg

• Bounding box and polygon annotation with interactive editing

• Keypoint annotation with COCO 17-point human pose and 5-point face templates; COCO preserves keypoints on export

• Auto-save mode for uninterrupted workflow

• Predefined class labels with customizable list

• Verification system to mark completed annotations

Workflow and Interface

Undo/Redo Support

Full undo/redo for annotation actions. Press Ctrl+Z to undo and Ctrl+Shift+Z to redo.

Balanced Modern Workspace

A fixed annotation-tool rail, compact command bar, unified Objects/Files inspector, integrated video timeline, and visible document status keep the active canvas uncluttered. Gallery Mode is embedded in the same workspace, so the annotation tools, inspector, and status remain available while browsing thumbnail previews across all five annotation formats.

• Colored borders indicate status: Gray (no labels), Blue (has labels), Green (verified)

• Bounding boxes use compact corner markers; polygons use outline previews

• Quick size presets (S/M/L/XL) plus slider for fine control

• Smart selection: click on nested boxes selects the inner box

• Press Ctrl+G to toggle gallery mode

Modern UI with Feather Icons

Official Feather icons provide a consistent, accessible tool vocabulary.

Responsive DPI Scaling

Icons and UI elements scale properly on high-DPI displays (4K, Retina).

Unified Object Inspector

Search and edit rectangles, polygons, keypoints, and video tracks in one Objects view. The Files view keeps the compact file list, thumbnails, and annotation-status filter close at hand, and the inspector can be collapsed when maximum canvas space is needed.

Faster Class Confirmation

New boxes, polygons, Smart Select results, and video geometry remain provisional while a non-modal class picker opens beside the annotation. Filter or enter a class, press Enter to commit one undoable annotation, or press Escape to discard it without changing the document. Default labels and established single-class sessions bypass the picker.

Smart Select Output Modes

While Smart Select is active, choose Box or Polygon from the compact canvas control. The choice persists between sessions, and either result follows the same provisional class-confirmation and undo workflow.

Portable Whole-Video Propagation

From the integrated timeline, propagate every accepted manual anchor on the current frame or only the selected object. The portable OpenCV backend processes rectangles, polygons, and associated keypoints together, shows preview-only progress, protects later manual anchors, and commits accepted observations and explicit gap records atomically in one undo step. Editing generated geometry creates a manual correction first, then regenerates only the bounded neighboring segments as a separate undoable change.

Optional SAM 2 Video Backend

Whole-video propagation can use an official, source-installed SAM 2 on Linux/CUDA. Auto is the default selection, and OpenCV remains the portable fallback. Open Tools → SAM Settings… to choose Auto, OpenCV, or SAM 2 and select a local checkpoint and its matching config file. Auto selects SAM 2 only with Python 3.10+, compatible CUDA-enabled PyTorch/torchvision and SAM 2, and both valid files; otherwise it uses OpenCV. An explicitly selected but unavailable SAM 2 reports the missing requirement and never silently falls back. labelImg++ does not download or bundle Torch, SAM 2, checkpoints, or configs, and none are added to its optional extras.

Direct Ultralytics Dataset Export

Choose Tools → Export Ultralytics Dataset… to create a ready-to-train YOLO detection dataset with images/{train,val,test}, matching labels/{train,val,test}, and data.yaml. Configure deterministic split ratios and either copy images or create absolute local symlinks. The destination must be new or empty, and it is published only after the full export succeeds. See the Ultralytics Export Guide.

Installed Python Plugins

Add trusted command plugins from separately installed Python distributions without editing labelImg++ source. Review and enable them under Tools → Plugins…; changes take effect after restart. Plugins use a versioned public API with host-owned actions, namespaced shortcuts and settings, bounded background work, read-only document state, diagnostics, and LABELIMGPP_DISABLE_PLUGINS=1 recovery. See the Plugin Authoring Guide.

Brightness Adjustment

Adjust image brightness on-the-fly to better see annotations on dark or light images.

Dark Mode Theme

Choose between light and dark themes for comfortable annotation in any lighting condition.

• Press Ctrl+Shift+T to toggle between themes

• Theme preference automatically saved

• All UI components (canvas, gallery, dialogs) respect the active theme

• See Dark Mode Documentation for detailed information

SAM-Assisted Segmentation (optional)

Click once on an object to auto-generate a polygon, traced from a Segment-Anything mask. Install the optional extra and toggle SAM Segment:

pip install labelimgplusplus[sam]

Runs the lightweight MobileSAM model on ONNX Runtime as a CPU-friendly extra with no PyTorch dependency. The default model pair is downloaded and checksum-verified on first use. Point Tools → SAM Settings… at your own exported encoder/decoder pair to use a different SAM variant. Without the extra installed, the action stays disabled (with an install hint) and nothing else changes.

Smart Video Annotation (optional)

Open local MP4, MOV, MKV, and AVI video, annotate on exact presentation timestamps, interpolate keyframes, and propagate rectangles, polygons, and associated keypoints across the video. Propagation streams preview-only results and commits accepted observations and gaps atomically; the legacy optical-flow suggestion workflow remains available separately. Video work is stored in a sibling <video>.labelimgpp.sqlite project; existing image annotations and formats are unchanged.

pip install labelimgplusplus[video]

The timeline supports frame stepping, exact timecode seeks, variable-rate media, playback without audio, track markers, and verified-frame markers. Accepted frames export to VOC, YOLO, YOLO-seg, COCO, or CreateML. See the Smart Video Annotation Guide for propagation setup, the project workflow, and export behavior.

From PyPI (Recommended)

Note: The command was renamed from labelImgPlusPlus to labelimgpp in v2.1.1. The old command still works but shows a deprecation warning.

pip3 install labelimgplusplus
labelimgpp

Or use the full command name:

labelimgplusplus

With a specific image or directory:

labelimgpp [IMAGE_PATH] [PRE-DEFINED CLASS FILE] [SAVE_DIR]

The first positional path can also be a local video or a .labelimgpp.sqlite video project. Optional features can be combined:

pip install labelimgplusplus[sam,video]

Build from Source

Ubuntu/Linux:

sudo apt-get install pyqt5-dev-tools
pip3 install -r requirements/requirements-linux-python3.txt
make qt5py3
python3 labelImgPlusPlus.py

macOS:

pip3 install pyqt5 lxml
make qt5py3
python3 labelImgPlusPlus.py

Windows:

pip install pyqt5 lxml
pyrcc5 -o libs/resources.py resources.qrc
python labelImgPlusPlus.py