allow convert_to_coco_dict for custom dataset without annotations

[wangg12]

Fix #1684 . Allowing inference when annotations are not available for custom datasets.

[ppwwyyxx]

Please describe the problem first. Currently, the issue is empty.

[wangg12]

I think the issue already describes the problem clearly:)

[ppwwyyxx]

An easier solution is to let convert_to_coco_json support dataset without annotations (instead of throwing an error). That should be doable with <5 lines.

[wangg12]

Maybe using

anns_per_image = image_dict.get("annotations", [])

in https://github.com/facebookresearch/detectron2/blob/master/detectron2/data/datasets/coco.py#L321.
I'm not sure if this is sufficient.

[wangg12]

I tested this, it works for me by just changing one line.

[ppwwyyxx]

Not sure whether

detectron2/detectron2/data/datasets/coco.py

Line 398 in e45cd68

"annotations": coco_annotations,

needs to be changed as well, given this line in evaluator:

detectron2/detectron2/evaluation/coco_evaluation.py

Line 86 in e45cd68

self._do_evaluation = "annotations" in self._coco_api.dataset

[wangg12]

By doing this, the evaluation will still be performed. But got all nan results.

[wangg12]

Now it should be fine. If there are no annotations at all, the converted coco json file will not contain "annotations".

[ppwwyyxx]

Looks great. Thanks!

[facebook-github-bot]

@ppwwyyxx is landing this pull request. If you are a Facebook employee, you can view this diff on Phabricator.

[facebook-github-bot]

@ppwwyyxx merged this pull request in 0f09111.