tlc.integration.detectron2.register_coco_instances¶

A drop-in replacement for detectron2.data.datasets.register_coco_instances

Module Contents¶

Functions¶

Function

Description

register_coco_instances

Register a COCO dataset in Detectron2’s standard format.

API¶

register_coco_instances(
name: str,
metadata: dict,
json_file: str,
image_root: str | None,
revision_url: str = '',
project_name: str = '',
keep_crowd_annotations: bool = True,
task: Literal[detect, segment] = 'detect',
mask_format: Literal[bitmask, polygon] = 'polygon',
) → None¶

Register a COCO dataset in Detectron2’s standard format.

This method works as a drop-in replacement for detectron2.data.datasets.register_coco_instances.

References:

The original function reads the json file and uses pycocoapi to construct a list of dicts which are then registered under the key name in detectron’s DatasetCatalog.

These dicts have the following format:

{
    "file_name": "COCO_train2014_000000000009.jpg",
    "height": 480,
    "width": 640,
    "image_id": 9,
    "annotations": [
        {
            "bbox": [97.84, 12.43, 424.93, 407.73],
            "bbox_mode": 1,
            "category_id": 16,
            "iscrowd": 0,
            "segmentation": [[...]]
        },
        ...
    ]
}

This function also registers a list of dicts under the key name in detectron’s DatasetCatalog, but before the data is generated, a Table is resolved. The first time the function is called with a given signature, a 3LC Table is created. On subsequent calls, the Table is replaced with the most recent descendant of the root table. If the resolved table contains a sample weight column, its value is sent along in the list of dicts under the key weight.

Parameters:

  • name – the name that identifies a dataset, e.g. “coco_2014_train”.

  • metadata – extra metadata associated with this dataset.

  • json_file – path to the json instance annotation file.

  • image_root – directory which contains all the images. None if the file_name contains a complete path.

  • revision_url – url to a specific revision of the table. If not provided, the latest revision will be used. If the revision is not a descendant of the initial table, an error will be raised.

  • project_name – the name of the project.

  • keep_crowd_annotations – Whether to keep annotations with iscrowd=1.

  • task – the task to register the dataset for.

  • mask_format – the format to use for the masks (only used when task is “segment”). Corresponds to the detectron2 config INPUT.MASK_FORMAT.

Returns:

None