Dataset export as Ultralytics-YOLO-Segmentation-1.0 doesn't respect annotation grouping.

[Abdul-Mukit]

Actions before raising this issue

• I searched the existing issues and did not find anything similar.
• I read/searched the docs

Steps to Reproduce

Situation:

1. Multiple polygons are grouped together during annotation. Keyboard shortcut: g.
2. Export the dataset as Ultralytics-YOLO-Segmentation-1.0 format.

Current Behaviour:
Grouping is not respected in the exported .txt label file. Each polygon, irrespective of its grouping, is exported as a separate annotation.

Expected Behavior

All polygons from the same group have to be joined together, using the function merge_multi_segment: https://github.com/ultralytics/ultralytics/blob/9c927162166247db178539e0c9287775090bbcef/ultralytics/data/converter.py#L529

A group of polygons should be exported as one annotation/polygon in the .txt label file.

Possible Solution

Adopt merge_multi_segment: https://github.com/ultralytics/ultralytics/blob/9c927162166247db178539e0c9287775090bbcef/ultralytics/data/converter.py#L529

Context

No response

Environment

CVAT Server version: 2.44.2
CVAT UI version: 2.44.2

[Moetez-Fradi]

Hi, I am new to open source contributing and I’d like to work on fixing this issue. My plan is to modify the export so that grouped polygons are merged using merge_multi_segment, and add tests to make sure grouped and ungrouped annotations export correctly. I will start immediately

[zhiltsov-max]

Hi, thank you for raising the issue. Do you have any information on how such grouped polygons should be represented in the exported file? From what I can see, the function returns a list of polygons.

[Moetez-Fradi]

Hello I'm back !
After deep investigations, I found that the error originates from the datumaro library during YOLO segmentation export. Specifically, it occurs inside the _export_yolo_ultralytics_segmentation function in cvat/apps/dataset_manager/formats/yolo.py function, when StreamDataset and its transformations (EllipsesToMasks, masks_to_polygons) are used. The issue is not in this code base, but in the Datumaro components handling the dataset and annotations. I checked the responsible class found at https://github.com/open-edge-platform/datumaro/blob/develop/src/datumaro/plugins/data_formats/yolo/exporter.py, and I found it doesn't take object groups into consideration. So I will raise the issue to the Datumaro repo.

[zhiltsov-max]

The upstream Datumaro does not support this format, it's implemented only in our fork: https://github.com/cvat-ai/datumaro. But in any case, if you know how the grouped polygons should be represented in the exported file, please add such information.

[Moetez-Fradi]

Grouped polygons in CVAT represent a single object split into multiple segments, which Ultralytics expects to be encoded as one YOLO-Segmentation line by merging all segments with merge_multi_segment at https://github.com/ultralytics/ultralytics/blob/9c927162166247db178539e0c9287775090bbcef/ultralytics/data/converter.py#L529

I will implement this in CVAT’s Ultralytics YOLO Segmentation exporter by grouping Polygon annotations by (label, group) and applying merge_multi_segment before writing the .txt files, and then open a PR with tests to validate grouped vs ungrouped behavior. I will let you know when I'm done !