Product Updates

CVAT Digest, May 2025: Smarter Data Tracking, Improved Data Import and Export, and an Expanded Analytics Suite Over the past month the CVAT team has focused on three themes: richer analytics, clearer control over data volume, and a cleaner, more predictable API. Below you’ll find an overview of the most noticeable changes along with a closer look at improvements that make daily annotation work faster and more reliable. Improved Analytics The most significant change you’ll notice lives in the new Analytics report, available from the “Actions” menu on every project, task, and job. The report now includes three tabs, each helping you to understand project progress better: Summary tab assembles high-level charts^total objects, time spent, average annotation speed, as well as a breakdown of shapes per label. Annotations tab lists every label alongside counts for polygons and other shapes, all searchable and filterable so large projects remain manageable. Events tab brings together log-level detail: task and job IDs, event type, frame count, object count, assignee, and timestamp. Because search and filter tools are built into each view, you can move from a bird’s-eye perspective to a single mislabeled frame without exporting data If you need to share findings, you can export data with all the filters applied, so no need to clean up spreadsheets. 3D Annotation (Online & On-Prem) We’ve added a quick way to annotate 3D objects with cuboids when object sizes are known in advance. Annotators can now create one (or several) correctly sized objects and simply copy/paste them into place, saving time and ensuring consistency. Automatic Data-Size Tracking CVAT now keeps an eye on the storage footprint of every image, video, and guide file you upload. For cloud users the measurement happens in the background; on self-hosted installations an administrator can initialize the scan with a single python manage.py initcontentsize command. Continuous tracking means you’ll see the true scale of a project before downloads, migrations, or backup operations begin. Improved Import and Better Export File-based annotation imports now tag their source as “file” by default, leaving no ambiguity. On the export side, event cache files move into a dedicated /data/cache/export/ directory, where they follow an automatic cleanup schedule so stale archives don’t accumulate. An API That Gets Out of the Way Several endpoints have been retired or consolidated to reduce duplication: Event logs now can be exported through POST /api/events/export, and the results can be fetched with a GET request to /api/requests/{rq_id}. Old endpoints were deprecated. Status checks for background requests were moved to the /api/requests/{rq_id} path, replacing the older quality-report endpoint. On the SDK side, classes such as DatasetWriteRequest and TaskAnnotationsUpdateRequest have been removed, and failed background tasks now raise a dedicated BackgroundRequestException rather than a catch-all ApiException. Project Management and Project Quality Settings Project-level quality settings now propagate automatically to every task under the project unless a task has its own custom rules. The Project Quality page has been tightened up so that all tasks render consistently, and a new job filter helps large teams isolate outliers without digging through spreadsheets. Orientation checks now save properly, so reviewers can apply the setting once and move on. Faster Search and Lower Memory Imports During annotation you can jump directly to a frame by typing part of the file name, which is handy for long video sequences. Behind the scenes, YOLO and COCO imports have been refactored to lower peak memory use, which in practice allows larger datasets to load without timeouts. Ultralytics YOLO archives import correctly even when image information is absent, and the Datumaro engine now supports ellipse shapes, improving a workflow for circular objects. Security and Reliability Fixes A browsable-API issue that exposed certain resource names and IDs has been closed. Restored backups now preserve the original asset owner, AI model tracking auto-starts as intended, and several corner-case crashes involving 3D cuboids and track interpolation have been addressed. Looking Ahead All the features described here are already available on CVAT Online and in the latest CVAT On-Premises release. As always, feedback is very valuable and drives our roadmap; if you have suggestions or run into friction, let us know through the usual channels. Until next month! You can read the full changelog here: https://github.com/cvat-ai/cvat/releases