AI-assisted dataset annotation skill that accepts frame and detection events via JSONL, supports bbox, SAM2, and DINOv3 annotation methods, and exports labeled datasets in COCO, YOLO, or VOC format.
Tasks
Run AI-assisted annotation on frames using the selected method (bbox, sam2, or dinov3) with auto-detection enabled
Review auto-detected bounding boxes, correct them, and add labels before exporting the dataset
Inputs
frame events containing camera_id, frame_path, frame_number, width, and height
detections events containing frame_number and a detections array with class, bbox, confidence, and track_id
save_dataset events containing a dataset name and export format
Outputs
ready events listing available methods and export formats
annotation events containing frame_number and an annotations array with category, bbox, track_id, and is_keyframe
dataset_saved events containing format, path, and stats with image, annotation, and category counts
Limitations and checks
The annotation protocol operates on individual frame events and bounding boxes; no mask or segmentation data is exchanged in the defined JSONL protocol
Object tracking is represented only via track_id and is_keyframe fields in annotation events; no interpolation mechanism is defined in the protocol
Export format options are coco, yolo, and voc, with coco as the default
Dataset upload to external platforms is mentioned in prose but has no corresponding input event, output event, or parameter in the protocol
Verify that the selected annotation method is one of bbox, sam2, or dinov3
Confirm that the export format is one of coco, yolo, or voc before triggering a save_dataset event
Ensure each annotation output includes a category, bbox, and track_id field
Check that dataset_saved output stats include counts for images, annotations, and categories