The fastest way to annotate computer vision datasets

Yes. Platform detects YOLO, COCO, LabelMe JSON, Labelbox NDJSON, Ultralytics NDJSON and semantic PNG masks automatically. You can import datasets from Roboflow with your API key, and bring in CVAT or Label Studio projects by uploading a YOLO or COCO export. Pascal VOC XML labels aren't imported, so convert them first.

Yes. Datasets export as Ultralytics NDJSON, except On Premise datasets. Image links in the export expire after 7 days, so use or convert it promptly. You usually don't need to convert the file: pass it or the dataset's ul:// URI to model.train() and Ultralytics builds the YOLO folders for you. To get YOLO folders without training, convert it with convert_ndjson_to_yolo in the ultralytics package.

Manual annotation means drawing each label yourself. Smart annotation proposes labels for you to review: click an object and a Segment Anything (SAM) model draws it, run a YOLO model (including your own) to pre-label, or let a class-prompted model find your classes on detection datasets with 1–200 classes. It works on every plan, for datasets stored on Platform rather than connected storage or On Premise. SAM clicks and hosted models are free; paid provider models bill your provider account.

Image annotation is the process of labeling images to identify objects, features, or regions within them. It is the foundational step in training computer vision models for tasks like object detection, image segmentation, image classification, and pose estimation. Annotation types vary by use case and include bounding boxes, polygons, masks, and keypoints. It's a process carried out in both open-source tools and dedicated commercial platforms.

Yes. Batch Annotation runs one model over your dataset: a YOLO model, including your own, or a class-prompted model on detection datasets with 1–200 classes. By default it labels only unlabeled images and saves a dataset version first, so you can roll back. It costs $1 per 1,000 images (minimum $0.01 per run); paid provider models also bill your provider account. It works on datasets stored on Platform, not connected storage or On Premise, and not on depth datasets.

Private datasets are available on every plan, but new datasets default to public on Free and Pro and private on Enterprise, so check the visibility toggle before uploading. Ultralytics Platform is ISO 27001:2022 certified and has a SOC 2 Type I report, with US, EU and APAC data regions. Enterprise On Premise keeps your images on your own computer for annotation and training; labels, metrics and trained models are stored in Platform.

Yes. Upload a video and Platform extracts frames for you to label as images, at up to 1 frame per second and at most 100 frames per video; longer videos are sampled at wider intervals. The frames are labeled individually, not as a continuous video sequence.