Auto-Label#
Annotating an image or video dataset by hand is tedious and time-consuming, and this is especially true for segmentation and reasoning tasks. Drawing an accurate polygon around an object can take 10 times longer than drawing a bounding box, and authoring question-answer pairs for a video can take longer still. The Auto-Label service of TAO Data Services reduces this effort by generating annotations automatically from existing images, videos, captions, and bounding boxes.
The Auto-Label service groups several pipelines under a single auto_label generate
command. You select a pipeline through the autolabel_type field in the
specification file, and each pipeline reads its own configuration block. The service
supports the following pipelines:
Mask Auto Label (MAL) generates instance segmentation masks from groundtruth bounding boxes.
Grounding DINO generates bounding box annotations from category names or noun chunks through iterative open-vocabulary detection.
2D Grounding uses a vision-language model to extract referring expressions from images and ground them to bounding boxes.
Video Reasoning Annotation uses vision-language and language models to produce reasoning-style question-answer annotations from videos.
Tip
Most pipelines also ship as a TAO skill that an agent runs for you, which is the
quickest way to get started. Each pipeline page opens with a “Quickstart with a TAO
Skill” section; the configuration tables and the auto_label generate command-line
path follow for when you need finer control.
Data Input for Auto-Label#
Each pipeline accepts a different input modality. The image-mask and open-vocabulary detection pipelines expect the groundtruth annotations for a directory of images in a COCO-format JSON file. The 2D Grounding pipelines accept image directories with captions or KITTI-format labels, and the Video Reasoning Annotation pipeline accepts a directory of videos or a list of JSONL files. Refer to each pipeline page for the exact input layout that it requires.
Running the Auto-Label Tool#
The Auto-Label service exposes a single task:
generate: Runs the pipeline selected byautolabel_typeand writes the generated annotations toresults_dir.
To run a pipeline, pass an experiment specification file to the generate task:
auto_label generate -e /path/to/spec.yaml results_dir=/path/to/results
Required arguments:
-e: Path to the experiment specification file.results_dir: Directory in which to write the generated annotations.
Refer to each pipeline page for the full configuration reference and a sample specification file.