Segment Anything#

Format specification#

SA-1B format specification is available here.

The SA-1B is a dataset consisting of 1 billion segmentation masks model generated. The model is known as SAM from FAIR and it can produce reliable and accurate segmentation masks given user prompts, such as points, boxes, or natural sentences.

Supported media types:

Image

Supported annotation types:

Bbox
Mask
Polygon
Ellipse

Supported annotation attributes:

predicted_iou
stability_score
crop_box
point_coords

Convert SA-1B dataset#

The SA-1B dataset is available here

A SA-1B dataset can be converted in the following way:

datum convert --input-format segment_anything --input-path <path/to/dataset> \
    --output-format <desired_format> --output-dir <output/dir>

An SA-1B dataset directory should have the following structure:

└─ Dataset/
    ├── <name1.ext>
    ├── <name1>.json
    ├── <name2.ext>
    ├── <name2>.json
    └── ...

If your dataset is not following the above directory structure, it cannot detect and convert your dataset as the SA-1B format properly.

Export to other formats#

It can convert the dataset into any other formats Datumaro supports. To get the expected result, convert the dataset to formats that support the specified task.

There are several ways to convert a SA-1B dataset to other dataset formats using CLI:

Convert a dataset from SA-1B format to COCO format:

datum convert --input-format segment_anything --input-path <path/to/dataset> \
    --output-format coco --output-dir <output/dir>

Or, using Python API:

import datumaro as dm

dataset = dm.Dataset.import_from('<path/to/dataset>', 'segment_anything')
dataset.export('save_dir', 'coco', save_media=True)

Export to SA-1B#

Please note that exporting to SA-1B format would drop label information in annotations due to the nature of the format.

There are several ways to convert a dataset to Segment Anything format:

Convert a dataset from COCO format to Segment Anything format:

# converting to segment_anything format from other format
datum convert --input-format coco --input-path <path/to/dataset> \
    --output-format segment_anything --output-dir <output/dir> -- --save-media

Examples#

Examples of using this format from the code can be found in the format tests