Update Datumaro dependency to 0.2.0 #3813
Conversation
sizov-kirill
changed the title
| dataset = Dataset.from_extractors(GetCVATDataExtractor( | ||
| instance_data=instance_data, include_images=save_images), env=dm_env) | ||
| with TemporaryDirectory() as tmp_dir: | ||
| dataset.export(tmp_dir, 'datumaro', save_images=save_images) |
There was a problem hiding this comment.
Please check if the annotations do not include absolute image paths.
What about the format tests?
Also, think about loading tracks, if all the annotations have a track_id attribute.
There was a problem hiding this comment.
Committed solution for problem with absolute image paths.
The format is tested along with all other formats in files: test_formats.py, test_rest_api_formats.py.
About tracks, I have a few thoughts:
- Not to support tracks with Datumaro format until Datumaro as a tool starts to support them.
- Or we can check for
track_idattribute for all formats in theimport_dm_annotationsfunction (in other case we'll just duplicate similar code as for MOT format).
How do you think which solution is better?
There was a problem hiding this comment.
The second solution looks better to me, but implementing it straight will cause unexpected side-effects. I think, we can try to add a common import option to import tracks, if possible. Not all formats will support this, of course, but those, that do, will utilize it. Not in this PR, though. MOT code is already duplicated with some adjustments in MOTS, btw, so I think the generic variant might be a bit complicated.
|
|
||
| class DeleteImagePath(ItemTransform): | ||
| def transform_item(self, item): | ||
| if not item.has_image: |
Motivation and context
How has this been tested?
Checklist
developbranch- [ ] I have increased versions of npm packages if it is necessary (cvat-canvas,cvat-core, cvat-data and cvat-ui)
License
Feel free to contact the maintainers if that's a concern.