Supervisely

Enables teams to annotate images using multiple geometric primitives (rectangles, polygons, skeletons, 3D lasso) with real-time collaboration, permission-based access control, and integrated AI models (SAM2, ClickSEG) that auto-generate annotations which annotators refine. The platform manages annotation state across concurrent users, tracks changes via audit logs, and enforces quality gates through review workflows before data enters training pipelines.

Provides frame-by-frame and track-based annotation for video sequences with automatic object tracking across frames, off-screen detection marking, and multi-view synchronization for multi-camera footage. The system maintains temporal consistency by propagating annotations forward/backward and detecting tracking breaks, allowing annotators to correct trajectories in bulk rather than per-frame. Supports pre-recorded video with on-the-fly transcoding (requires Video Max add-on) and CDN acceleration for large files.

Generates synthetic training data by applying transformations (rotation, scaling, color jittering, blur) to existing annotations, or by rendering 3D models in simulated environments. Supports both image-level augmentation (modify existing images) and scene-level synthesis (render new scenes from 3D assets). Generated data is versioned and tracked separately from human-annotated data. Integration with model training allows teams to augment datasets on-the-fly during training.

Provides a training orchestration layer that manages model training runs, hyperparameter tuning, and result tracking. Supports integration with popular frameworks (PyTorch, TensorFlow — unclear if both are supported) and custom training scripts. Training runs are logged with dataset version, hyperparameters, metrics, and model weights. Results are compared across runs to identify best-performing models. Hardware specifications for training (GPU type, memory, timeout) are unknown.

Provides search capabilities across images, annotations, and metadata using both keyword search (filename, class name) and semantic search (find similar images based on visual content). Supports filtering by annotation properties (class, confidence, annotator, date), metadata tags, and custom attributes. Search results can be exported as new datasets or used to create subsets for targeted annotation or analysis. Semantic search uses embeddings (model unknown) to find visually similar images.

Enables multiple annotators to work on the same image simultaneously with real-time synchronization of changes. Detects conflicts when two annotators modify the same annotation and flags them for resolution. Supports undo/redo with conflict awareness (undo by one user doesn't affect another user's changes). Annotation state is persisted to the server after each change, ensuring no data loss. Latency and conflict resolution strategy are unknown.

Enables annotation of 3D point clouds (LiDAR, RADAR, depth sensors) with cuboid, cylinder, and segmentation primitives, with synchronized 2D image context from camera feeds to resolve ambiguities. The platform fuses multi-sensor data (e.g., LiDAR + camera + radar) into a unified 3D scene, allowing annotators to label objects in 3D space while referencing 2D projections. Includes automatic ground segmentation and AI-assisted cuboid generation (requires Cloud Points Max add-on at €399/month).

Provides specialized annotation tools for DICOM medical imagery including multi-planar reconstruction (MPR), 3D perspective views, and slice-by-slice segmentation with automatic 3D tracking across slices. Includes anonymization tools to strip PHI (patient identifiers, dates) and enforce HIPAA compliance. Medical Max add-on (€149/month) unlocks 50,000+ file limit, 3D tracking, and anonymization features. Supports CT, MRI, X-ray, and ultrasound modalities.

Provides one-click auto-labeling using pre-trained foundation models (YOLOv11 for detection, RT-DETRv2 for real-time detection, MM Segmentation for semantic segmentation, SAM2 for instance segmentation) directly integrated into the annotation UI. Annotators can trigger auto-labeling on single images or batch-process entire datasets, then refine predictions. Supports custom model integration via SDK for proprietary models; custom models run within Supervisely infrastructure (hardware specs unknown).

Enables definition of nested class hierarchies (e.g., Vehicle > Car > Sedan) and arbitrary key-value attributes (color: red, damaged: true) attached to annotations. The ontology is enforced at annotation time, preventing invalid class combinations and ensuring consistent labeling across teams. Supports conditional attributes (e.g., 'license plate' attribute only appears if class is 'car'). Ontologies are versioned and can be updated retroactively with migration rules.

Implements multi-stage QA workflows where annotations are reviewed by senior annotators or QA specialists before entering the training dataset. Supports consensus-based review (multiple annotators label same image, system flags disagreements) and conflict resolution (reviewer adjudicates between conflicting annotations). Tracks QA metrics (agreement rate, reviewer corrections) and can reject annotations below quality thresholds. Integrates with permission system to enforce role-based access (annotator vs. reviewer).

Tracks dataset versions as annotations are added, modified, or removed, allowing teams to reproduce model training runs with specific dataset snapshots. Each version captures metadata (number of images, classes, annotators, QA status) and can be tagged with experiment identifiers. Supports branching (create alternate versions for A/B testing) and merging (combine annotations from multiple branches). Integrates with model training to log which dataset version was used for each training run.

Provides Python SDK and AppEngine for building custom labeling applications that extend the platform's built-in tools. Developers can create custom UIs, integrate proprietary models, or build domain-specific workflows (e.g., medical image registration, 3D reconstruction). Apps are deployed to Supervisely infrastructure and appear as buttons in the annotation UI. SDK includes bindings for data access, model inference, and UI components. No documentation on language support, versioning, or deployment process.

Organizes datasets into hierarchical project structures (Workspace > Project > Dataset) with role-based access control (Owner, Manager, Annotator, Reviewer, Viewer). Permissions cascade down the hierarchy; granting access to a Workspace grants access to all Projects within it. Supports team management (invite users, assign roles, revoke access) and audit logging of all permission changes. Integrates with annotation tools to enforce permissions (e.g., Annotators cannot delete annotations).