Fix inconsistent force_masks across dataset formats

[kirilllzaitsev]

Description

load_coco_annotations used force_masks as a flag that indicates if masks should be parsed or not. On the contrary, YOLO and VOC enforce parsing masks if this flag is set, or parse masks in case they are present in annotations. Current change homogenizes the behavior, setting it to the second option.

Type of change

Please delete options that are not relevant.

• [+] Bug fix (non-breaking change which fixes an issue)

How has this change been tested, please provide a testcase or example of how you tested the change?

Does not impact anything.

Any specific deployment considerations

For example, documentation changes, usability, usage/costs, secrets, etc.

Docs

• Docs updated? What were the changes:

[SkalskiP]

Hi, @kirilllzaitsev! 👋🏻 I marked this PR as a 0.14.0 release. I'm super busy with wrapping work on a few 0.13.0 features as we plan to release tomorrow. Please be patient 🙏🏻

@hardikdava and @artyaltanzaya could you take a look at the PR in the meantime?

[SkalskiP]

@capjamesg @yeldarby I think you should know we are working on those changes. They might influence Autodistill.

[CLAassistant]

All committers have signed the CLA.

[SkalskiP]

Hi, @kirilllzaitsev 👋🏻! Sorry for not being overly active on this PR. I was a bit busy. Any chance you could provide Google Colab to test the updated version? 🙏🏻

[kirilllzaitsev]

@SkalskiP sure, I will prepare it a bit later today

[SkalskiP]

Awesome! Thanks a lot @kirilllzaitsev 🙏🏻

[artyaltanzaya]

thanks for the contribution! added minor suggestions for clarity.

[kirilllzaitsev]

@SkalskiP here is the colab link. Current behavior in COCO and VOC differs if the force_masks is True, but there are no masks: VOC proceeds with an empty annotation.mask, while COCO errors out with a KeyError (at this point, image_annotation["segmentation"]).

I would handle the case of force_masks=True and no masks with a warning and no error. What are your thoughts?

[SkalskiP]

Hi, @kirilllzaitsev 👋🏻! Will you have time to resolve conflicts in this branch today?

[kirilllzaitsev]

Hey @SkalskiP , sure, done

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 74%. Comparing base (b5ccf68) to head (c0fbac0).
⚠️ Report is 1 commits behind head on develop.

Additional details and impacted files

@@          Coverage Diff           @@
##           develop   #265   +/-   ##
======================================
  Coverage       73%    74%           
======================================
  Files           61     61           
  Lines         7390   7395    +5     
======================================
+ Hits          5422   5447   +25     
+ Misses        1968   1948   -20     

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.

[Copilot]

Pull request overview

This PR aims to make force_masks behave consistently across dataset formats by enabling masks when the flag is set or when masks are present in the annotations, instead of using force_masks as a strict on/off switch for parsing masks in COCO.

Changes:

• In pascal_voc.py, introduces a _with_mask helper and updates detections_from_xml_obj to use it when deciding whether to include masks.
• In coco.py, updates load_coco_annotations to compute with_masks based on force_masks or the presence of a segmentation field and passes this into coco_annotations_to_detections, adding a _with_mask helper for COCO annotations.

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 2 comments.

File	Description
supervision/dataset/formats/pascal_voc.py	Refactors mask detection into a _with_mask helper and adjusts detections_from_xml_obj mask-flag logic to align with other formats.
supervision/dataset/formats/coco.py	Changes load_coco_annotations to infer with_masks from force_masks plus annotation content via a new _with_mask helper and passes the result into detection conversion.

Comments suppressed due to low confidence (2)

supervision/dataset/formats/pascal_voc.py:245

• The with_masks flag is overwritten on each object inside the loop and then only used once when constructing Detections, so if an earlier object has a polygon and a later object does not, previously computed masks will be discarded because with_masks reflects only the last object processed. To make mask handling consistent and avoid silently dropping masks, consider computing with_masks based on whether any object in the image has a polygon (or on force_masks) outside the loop, or otherwise ensuring it accumulates mask presence across objects rather than just mirroring the last one.

        with_masks = force_masks or _with_mask(obj)

        for polygon in obj.findall("polygon"):
            polygon = parse_polygon_points(polygon)
            # https://github.com/roboflow/supervision/issues/144

supervision/dataset/formats/pascal_voc.py:245

• detections_from_xml_obj has mask-handling logic controlled by force_masks and _with_mask, but the existing tests in test/dataset/formats/test_pascal_voc.py::test_detections_from_xml_obj only cover bounding boxes and never exercise objects with <polygon> elements or force_masks=True. Adding tests that cover polygon masks and the force_masks flag would help ensure that future changes to this code (including the new _with_mask helper) behave as intended.

        with_masks = force_masks or _with_mask(obj)

        for polygon in obj.findall("polygon"):
            polygon = parse_polygon_points(polygon)
            # https://github.com/roboflow/supervision/issues/144

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.

[Copilot]

Pull request overview

Copilot reviewed 5 out of 5 changed files in this pull request and generated 3 comments.

---

💡 Add Copilot custom instructions for smarter, more guided reviews. Learn how to get started.