When creating an image segmentation Dataset via the Python SDK (code below), the instance seems to always be of the type bitmask. From the UI it is possible to choose between "bitmask", "polygon" and "polyline" and when uploading a jsonl file containing polygon labels, this seems to work.

Code to create and upload image segmentation dataset via SDK:

import google.cloud.aiplatform as vertex_ai

vertex_ai.init(project=project, location=location)

dataset = vertex_ai.ImageDataset.create(
    display_name="test",
    gcs_source=gcs_jsonl,
    import_schema_uri=vertex_ai.schema.dataset.ioformat.image.image_segmentation,
    sync=sync,
)

With an example gcs_jsonl file:

{"imageGcsUri": "gs://image.jpg", "polygonAnnotations": [{"displayName": "potato-poly", "vertexes": [{"x": 0.000803125, "y": 0.9318263888888888}, {"x": 0.0267796875, "y": 0.9394652777777778}, {"x": 0.037521874999999996, "y": 0.9658541666666666}, {"x": 0.0392796875, "y": 1.0}, {"x": 0.0, "y": 1.0}]}]}

Note that there is no other image segmentation schema than the one that is used and that the format for polygons in the jsonl is correct (when that jsonl file is uploaded after an image segmentation dataset of type polygon is created via the UI, it works as it should).

It looks like the command vertex_ai.ImageDataset.create misses the option to define the type.

Conclusion:

From the SDK it is not possible to create an image segmentation dataset of type "polygon" or "polyline", the type "bitmask" is always used.

Issue 226434404

Links (4)

Issue metadata