| Welcome to the October edition of the CVAT Digest! |
| This month, we shipped four releases: v2.78.0, v2.79.0, v2.80.0, and v2.81.0. Highlights include: |
- Multi-object selection and batch editing.
- Annotation export for audio projects.
- More flexible replica job workflows.
- Mask and polygon tracking directly in the browser.
- Custom text prompts for SAM3 multi-object segmentation.
|
| |
| What shipped this month |
 | Feature Highlight Select and edit multiple objects at once You can now select several annotation objects and work with them together. Move, copy, paste, delete, or change labels in one operation, and copy selections between video frames while preserving their relative positions, labels, and attributes. Read more → |
|
| Track masks and polygons in your browser |
| The new Mask & Polygon Tracker follows existing annotations across video frames using a model that runs locally in your browser. No separate model server or AI agent is needed. Performance depends on your device and WebGPU support. |
| Available in CVAT Online paid plans and CVAT Enterprise. |
| Use custom text prompts with SAM 3 |
| SAM 3 now accepts custom text descriptions for multi-object segmentation. Find matching objects using text alone or combine text with visual prompts, without changing your project labels. |
| Available in CVAT Online paid plans and CVAT Enterprise. |
| Export audio annotations as Generic TSV |
| Export labeled audio intervals and transcription data in a single Generic TSV 1.0 file, including timestamps, labels, attributes, and source filenames. Read docs → |
| More flexible replica jobs and quality feedback |
| Create and manage replica jobs separately from consensus merging. Immediate Job Feedback now highlights confused classes, missing annotations, and extra annotations after rejected validation attempts. Read docs → |
| Other improvements |
| October also brought better performance in annotation-heavy jobs, reorganized AI assistance controls, and refreshed subscription management and access notices in CVAT Online. |
| See the full changelog → |
| |
| From Community |
| Each month we spotlight interesting datasets, models and other projects created with the help of CVAT and publicly shared by the community. Working on something cool? Let us know! |
- Mapping kelp forests from satellite imagery 🌊: A team including the University of Victoria and Hakai Institute used CVAT to annotate 9,856 kelp polygons across ten Sentinel-2 scenes for automated coastal mapping.
- Drone-based pothole detection 🛣️: The MyUAV-UTHMPGR01 dataset contains 763 drone images of roads in Malaysia, manually annotated in CVAT to benchmark pothole detection models.
- Pose estimation for rehabilitation robotics 🦾: INESC TEC researchers used CVAT to annotate eleven upper-body keypoints and benchmark ten pose-estimation models in an upper-limb exoskeleton setup.
|
|
Support CVAT on GitHub ⭐
If CVAT helps your work, consider giving us a star on GitHub. It helps more researchers, developers, and teams discover the project.
|
|
 |
| |