fix: return empty int ndarray instead of None for class_id on empty VLM parse

[YousefZahran1]

What

On empty-parse paths in two VLM parsers, class_id was returned as None instead of a zero-length ndarray of dtype int, inconsistent with every other exit point in these functions and with the convention established by from_google_gemini_2_5, from_qwen_2_5_vl, and from_qwen_3_vl.

Affected code paths

Function	Condition	Before	After
from_paligemma	regex finds no matches	None	np.empty((0,), dtype=int)
from_google_gemini_2_0	JSONDecodeError	None	np.empty((0,), dtype=int)
from_google_gemini_2_0	empty bounding-box list	None	np.empty((0,), dtype=int)

Why it matters

Callers that do class_id.shape, len(class_id), or class_id[mask] immediately after parsing will get an AttributeError / TypeError whenever zero detections are returned — a very common path in production (frame with nothing to detect). Downstream Detections.from_vlm(...) pipelines hit this at runtime while unit tests may pass because they check happy paths.

Changes

• src/supervision/detection/vlm.py — 3 one-line fixes (no logic change)
• tests/detection/test_vlm.py — updated 10 test expectations to match the new contract

All 88 tests in tests/detection/test_vlm.py pass.

Fixes #2219

[CLAassistant]

All committers have signed the CLA.

[Borda]

Thank you for your submission! We really appreciate it. Like many open source projects, we ask that you sign our Contributor License Agreement before we can accept your contribution.

@YousefZahran1 could you pls sign ^^ 🦝

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.
✅ Project coverage is 78%. Comparing base (f4b0767) to head (8bb5f1d).

Additional details and impacted files

@@           Coverage Diff           @@
##           develop   #2239   +/-   ##
=======================================
  Coverage       78%     78%           
=======================================
  Files           66      66           
  Lines         8374    8374           
=======================================
  Hits          6501    6501           
  Misses        1873    1873           

🚀 New features to boost your workflow:

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

[YousefZahran1]

"CLA signed! Ready for review when you have a moment.

[Copilot]

Pull request overview

This PR standardizes the “empty parse” return contract for VLM parsers so class_id is always a zero-length np.ndarray of dtype=int (instead of None) when no detections are produced, aligning PaliGemma and Gemini 2.0 behavior with other parsers and preventing downstream AttributeError/TypeError on common empty-detection paths (see #2219).

Changes:

• Update from_paligemma(...) to return np.empty((0,), dtype=int) for class_id when regex matching yields zero detections.
• Update from_google_gemini_2_0(...) to return np.empty((0,), dtype=int) for class_id on JSON decode failure and when no boxes are parsed.
• Adjust unit test expectations for those empty-result scenarios.

Quality assessment (n/5):

• Code quality: 4/5
• Testing: 4/5
• Docs: 4/5

Reviewed changes

Copilot reviewed 2 out of 2 changed files in this pull request and generated 1 comment.

File	Description
src/supervision/detection/vlm.py	Align empty-result class_id to an empty int ndarray for PaliGemma and Gemini 2.0 error/empty paths.
tests/detection/test_vlm.py	Update expected tuples for empty-result cases to match the new class_id contract.

Comments suppressed due to low confidence (1)

src/supervision/detection/vlm.py:630

• In the classes is not None branch, class_id = np.array([...]) does not specify dtype=int, so if the class filter removes all detections it returns an empty float array. This undermines the new behavior of returning an empty int ndarray on other empty-result paths; consider forcing dtype=int for consistent downstream handling.

    try:
        data = json.loads(result)
    except json.JSONDecodeError:
        return np.empty((0, 4)), np.empty((0,), dtype=int), np.empty((0,), dtype=str)