Bounding Box Overlap Filtering

In object detection systems, machine learning models frequently generate multiple candidate bounding boxes for the same object. These overlapping predictions create redundancy that must be resolved before presenting results to end users or downstream systems.

Bounding Box Overlap Filtering is an essential post-processing algorithm used to eliminate duplicate detections by analyzing the spatial overlap between candidate boxes and their associated confidence scores. The algorithm systematically retains the most confident prediction for each detected object while suppressing boxes that significantly overlap with already-selected detections.

Algorithm Mechanism

The filtering process operates through the following iterative procedure:

1. Selection Phase: Identify the bounding box with the highest confidence score among all remaining candidates
2. Evaluation Phase: Compute the Intersection-over-Union (IoU) between the selected box and every remaining candidate
3. Suppression Phase: Eliminate all boxes whose IoU with the selected box exceeds the specified threshold
4. Iteration: Repeat until no candidate boxes remain

Intersection-over-Union (IoU) Metric

The IoU between two bounding boxes quantifies their spatial overlap and is defined as:

$$\text{IoU} = \frac{\text{Area of Intersection}}{\text{Area of Union}}$$

Where:

• Intersection Area = Area of the overlapping region between two boxes
• Union Area = Total area covered by both boxes combined (Intersection counted once)

IoU values range from 0 (no overlap) to 1 (perfect overlap).

Input Specification

• boxes: A 2D array of shape (N, 4) containing N bounding boxes. Each box is defined as [x1, y1, x2, y2] where:
  • (x1, y1) represents the top-left corner coordinates
  • (x2, y2) represents the bottom-right corner coordinates
• scores: A 1D array of shape (N,) containing confidence scores for each corresponding box
• iou_threshold: A floating-point value in range [0, 1] defining the overlap threshold for suppression

Output Specification

• Return a list of indices of the retained boxes, ordered by descending confidence score
• Return -1 for invalid inputs:
  • Mismatched dimensions between boxes and scores
  • Invalid box format (not 4 coordinates per box)
  • Threshold outside valid range [0, 1]
• Return an empty list for empty input arrays

Your Task

Implement the bounding box overlap filtering algorithm that efficiently processes detection candidates and returns the indices of non-redundant, high-confidence bounding boxes.