Matching Logic

This page outlines the unit test coverage for the matching.py script, which compares YOLO and Grounding DINO predictions and routes files into labeled or pending folders based on IoU, confidence, and class agreement.

Coverage Overview

This test suite validates the following core functions:

  • compute_iou: Intersection-over-Union between two bounding boxes

  • normalize_class: Standardizes class names (e.g., removes suffixes)

  • match_predictions: Compares YOLO and DINO predictions using class name and IoU

  • match_and_filter: Reads YOLO/DINO JSON predictions and splits them into labeled or pending output folders using thresholds

Unit Test Descriptions

compute_iou

  • test_compute_iou_overlap
    Validates IoU is correctly calculated for overlapping boxes.

  • test_compute_iou_no_overlap
    Ensures IoU returns 0 for non-overlapping boxes.

normalize_class

  • test_normalize_class_strips_suffix
    Checks normalization of class strings (e.g., lowercasing, suffix removal).

match_predictions

  • test_match_predictions_positive
    Confirms that two bounding boxes with high IoU and same class are matched.

  • test_match_predictions_negative_iou
    Confirms that boxes with low IoU do not match even if class names are the same.

match_and_filter

All tests below use the match_dirs fixture, which sets up temporary yolo, dino, labeled, and pending directories for testing.

  • test_match_and_filter_matched_goes_labeled
    Valid match (class + IoU + confidence) goes to labeled.

  • test_match_and_filter_low_conf_goes_pending
    Low confidence YOLO prediction results in pending output.

  • test_dino_false_negative_goes_pending
    High-confidence DINO detection missed by YOLO triggers pending file.

  • test_match_and_filter_invalid_yolo_json_skipped
    Handles bad JSON gracefully — file is skipped, not labeled or pending.

Configuration Parameters Used

These thresholds are passed as a dictionary to match_and_filter and influence labeling decisions:

Key

Description

iou_threshold

Minimum IoU for a match (e.g., 0.5)

low_conf_threshold

YOLO predictions below this are flagged as false positives

mid_conf_threshold

Intermediate YOLO confidence triggers review

dino_false_negative_threshold

DINO detections above this are considered missed by YOLO

Summary

This page outlines robust coverage of the matching.py logic, ensuring:

  • Reliable IoU calculation

  • Consistent class name normalization

  • Correct behavior across all matching outcomes

  • Fault tolerance for malformed files

  • Accurate routing of predictions based on configurable thresholds

Together, these tests ensure that only confident, well-aligned predictions are passed automatically to the labeled dataset, while edge cases and uncertainties are flagged for human review.