Welcome to the March edition of the CVAT product roundup. This month, we added snap-to-point and polygon join tools to make shape editing easier, introduced label-based text prompts for SAM 3 to speed up segmentation of repeating objects, and shipped a set of practical improvements across storage, exports, and infrastructure.

The updates in this roundup were shipped across v2.59.0, v2.59.1, v2.60.0, and v2.61.0.

What’s new

Snap to Point for Polygon and Polyline Editing‍

You can now use Snap to point when editing polygons and polylines. With this enabled, the point you’re drawing or moving will automatically align with nearby points from other shapes. This is especially useful in dense annotation tasks where shapes share borders or need to line up precisely. Instead of manually matching vertices one by one, annotators can place points faster and keep geometry more consistent across the dataset.

Join Tool Now Supports Polygons‍

You can now use the Join tool with polygons to merge multiple polygon shapes into one unified polygon. This is useful when a single object was initially labeled in separate parts or when an annotator needs to clean up segmentation work without redrawing the shape from scratch. It makes polygon editing more flexible and reduces manual rework in segmentation-heavy projects.

Use Existing Labels as Text Prompts for SAM 3

We’ve added a new class-guided segmentation workflow with SAM 3. You can now use the labels defined during project or task setup as text prompts, then click or box one example object as a visual prompt to let SAM 3 find and segment the other matching objects of that class in the same image. Once the results are generated, you can adjust the confidence threshold to control how selective the predictions are and refine the output if needed.

This is a major improvement for images with repeating objects, because it turns a one-by-one segmentation task into a much faster class-based workflow.

Learn more

Docker Compose for Running the SAM 2 Agent‍

We’ve added Docker Compose support for running the SAM2 agent. For teams testing or deploying model-assisted segmentation workflows, this lowers setup friction and makes it easier to get the agent running in a more reproducible environment.

Docker + Compose for YOLO‍

We’ve also added Docker and Compose support for YOLO. This should make it easier to stand up YOLO-based workflows and reduce environment-specific setup overhead for teams using model-assisted annotation or integrating detection pipelines into CVAT.

Cloud Storage as Backing Storage for Local Tasks‍

You can now use cloud storage as backing storage for local tasks. This gives teams more flexibility in how they manage data, especially when they want to keep task workflows simple in CVAT while still relying on cloud-backed storage underneath. For larger datasets or mixed storage setups, this can reduce unnecessary duplication and make storage management cleaner.

Download Projects, Tasks, and Jobs Lists as CSV‍

You can now export Projects, Tasks, and Jobs metadata as CSV. This is a practical addition for reporting, audits, internal tracking, or sharing operational data outside CVAT. Instead of manually copying information from the UI, teams can now pull structured lists directly for analysis or reporting workflows.

Learn more

Updated

AI Tooling Is Now More Consistent Under the Hood‍

The interface behind AI-assisted tools is now aligned more closely with the /annotations API and AI detector outputs, with the IoG serverless function updated accordingly. This makes integrations easier to maintain and helps reduce mismatch between different AI workflows in CVAT.

Parent and Replica Jobs Now Use a Flat List View‍

Parent and replica jobs now appear in a flat list with clear “parent” and “replica” tags. Replica jobs remain hidden by default through filtering, but can still be shown when needed. This makes the jobs view easier to scan and understand for teams using consensus or replica-based review workflows.

Filters Now Work Better with Browser Navigation‍

We’ve updated filters on task and jobs pages so they now create browser history entries. That means you can move Back and Forward between filter states more naturally.

More Consistent Tab Navigation in 2D and 3D Workspaces‍

Tab and Shift+Tab behavior is now consistent across 2D and 3D workspaces. For teams working across different data types, this makes navigation more predictable and reduces friction between workspaces.

Models Page Is Now Always Visible in the UI‍

The Models page now remains visible even if the serverless module is not installed. This makes the interface more consistent and helps you understand where model-related functionality lives, even in deployments where that module is not currently enabled.

Fixed

1. Improved string escaping in CSV export.
2. Resolved an issue with ground truth annotation locking in review mode for regular jobs.
3. Corrected the Location header returned by the task data upload endpoint when requests are made without a trailing slash.
4. Reduced front-end memory usage by replacing standard JavaScript arrays with typed arrays.
5. Fixed memory leaks related to unreleased object URLs in cvat-canvas.
6. Resolved an SDK crash affecting backup creation and chunk upload recovery.
7. Improved Helm support for modifiable permissionFix paths and commands.