BasicAI

The image tool covers bounding boxes, polygons, polylines, keypoints, and semantic segmentation. Video adds frame by frame tracking so one object keeps a single identity across a clip. Supported image formats run from JPEG and PNG to BMP, TIFF, and WebP. Speed on routine detection work is the point: track an object once, and the tool carries it through later frames for you to correct.

This is what separates BasicAI from general purpose labeling tools. Point cloud data and camera images share one interface, so an annotator can draw a 3D box and see it projected onto the photo, or start on the image and check depth in the cloud. Detection, tracking, and segmentation modes each have their own flow, and camera calibration parameters are managed inside the platform.

A dedicated dataset type handles language model training data. SFT mode pairs prompts with reference answers, while RLHF mode records comparative judgments, usually scores or rankings, that pull a model toward human preferences. Conversation level annotation covers instruction tuning and preference collection without a second text annotation vendor.

Pre trained models label new data before a human touches it. Your own model can join through the integration APIs: text classification, entity recognition, image detection, image segmentation, point cloud detection, and point cloud segmentation. On repetitive datasets the annotator's job shifts from drawing every box to checking and fixing, which is a large time saving.

QA rules run against data, objects, or attributes and can flag missing labels, stray objects, or items outside a defined valid area. Checks run two ways, live in the tool while someone annotates and as a bulk job across a whole dataset. Rules attach to the dataset, so task flows inherit them instead of retyping each batch.

Work splits by role. Project managers hold data and assign tasks, annotators claim and label, reviewers check submissions, and inspectors make the final acceptance call. Tasks can pass several review stages, and reports show project status, per task progress, and annotator performance. It's a structure for outsourced labeling, not for one researcher drawing boxes late at night.

The REST API covers dataset creation, task management, file upload, and downloading results, with keys scoped to a team and expiry dates you control. Files go up through pre signed URLs, which keeps large batches off the API itself. Default rate limits leave room for scripting, at 100 calls a second and 3,000 a minute.