[Improvement] Increase Unit Test Coverage to ~90%

[Borda]

📝 Description

As supervision matures into an industry-standard utility package for Computer Vision, the reliability and stability of the codebase become paramount. Currently, our total test coverage sits at approximately 50%, with critical modules like metrics and key_points falling well below 40%.

To ensure this package remains the go-to solution for production pipelines and research alike, I propose an initiative to systematically increase our unit test coverage to a realistic target of ~90%.

🚀 Motivation

Why is this high priority?

• Industry Standard Responsibility: supervision is infrastructure. Users rely on it for critical CV workflows. A subtle bug in a metric calculation (like mAP or confusion matrix) can silently invalidate research experiments.
• Refactoring Confidence: As we optimize performance, a robust test suite acts as a safety net.
• Handling Edge Cases: Computer Vision data is notoriously messy. High coverage ensures we handle empty detections, None values, and malformed shapes gracefully.

📊 Current State

According to the latest Codecov report, we have significant gaps in core modules.

Module	Coverage	Status	Notes
metrics	38.36%	🔴 Critical	Over 1,100 lines untested.
key_points	36.68%	🔴 Critical	Lowest percentage in the library.
utils	55.06%	🟡 Low
dataset	57.32%	🟡 Low	Formats logic needs hardening.
detection	68.49%	🟡 Moderate	High volume of missed lines (639) due to file size.

🛠 Proposed Strategy

We should tackle this modularly. The immediate priority is to lift the "Red" modules out of the danger zone.

Priority 1: The Criticals

• metrics/: This is the highest impact area. We need to verify that mathematical calculations for accuracy, recall, and precision are correct across various edge cases.
• key_points/: Needs basic functional tests for skeleton handling and visualization.

Priority 2: Stability Layers

• dataset/: Ensure saving/loading functions (YOLO, COCO, Pascal VOC) are robust against file system errors and malformed inputs.
• utils/: General utility functions that are used everywhere need to be bulletproof.

✅ Guidelines for New Tests

For contributors looking to help:

• Test Behavior, Not Implementation: Focus on inputs and expected outputs.
• Edge Cases: Explicitly test for empty lists, None values, and mismatched array shapes.
• Mocking: Use mocking for external I/O (downloading files, displaying windows) to keep the suite fast.
• Parametrization: Use @pytest.mark.parametrize to cover multiple scenarios efficiently.

📋 Tracking

Please comment below if you would like to pick up a specific module so we don't duplicate work.

• supervision/metrics (High Priority)
• supervision/key_points (High Priority)
• supervision/dataset
• supervision/utils
• supervision/detection

---

The actual coverage report can be found at https://app.codecov.io/gh/roboflow/supervision/blob/develop

[danielsamo-se]

I’ll add tests for the Detections.from_* adapters (YOLOv5 / Ultralytics / YOLO-NAS) in detection/core.py

[panagiotamoraiti]

Hi @Borda, I could take a look at the metrics for detection. I’ve already worked on a pull request for confusion matrix calculation, so that might be a good starting point.

[uahroberto]

Hi @Borda! 👋

I'd like to help improve coverage for supervision/metrics.

I see @panagiotamoraiti is focusing on the Confusion Matrix part. To avoid overlaps, I can tackle the core accuracy metrics (mAP, Recall, Precision) and general metric utilities.

Bonus: This aligns with my plan to refactor Strict Typing for metrics in issue #2088. I can combine both efforts (Typing + Unit Tests) to ensure the module is mathematically correct and strictly typed.

Let me know if this split works for you!

[Borda]

I can combine both efforts (Typing + Unit Tests) to ensure the module is mathematically correct and strictly typed.

That is perfects, let's just keep it as two separate contributions/PRs SOS we reduce potential collisions and make it smoother to land...