Principle:Cleanlab Cleanlab Object Detection Visualization
| Knowledge Sources | Cleanlab |
|---|---|
| Domains | Machine_Learning, Data_Quality, Object_Detection |
| Last Updated | 2026-02-09 |
Overview
Visual inspection technique for object detection label issues that renders annotated bounding boxes on images to compare ground truth labels against model predictions.
Description
Visualization overlays ground truth and predicted bounding boxes on the original image, using different colors to distinguish label types and highlight discrepancies. This enables human reviewers to visually verify detected label issues and understand the nature of each error (overlooked, swapped, or badly located).
The visualization supports two display modes:
- Overlay mode (default): Both ground truth and prediction boxes are rendered on a single image, allowing direct visual comparison.
- Side-by-side mode: Ground truth and predictions are shown on separate copies of the image for clearer distinction when boxes overlap significantly.
Each bounding box is annotated with its class name (when class names are provided) and color-coded to distinguish ground truth from predictions. This makes it straightforward to spot:
- Overlooked objects: Prediction boxes with no corresponding ground truth box.
- Swapped labels: Matched boxes where the class labels differ.
- Badly located boxes: Matched boxes where the positions are significantly different.
Usage
Object detection visualization is used as the final step in a label quality audit workflow. After computing quality scores and identifying images with label issues, reviewers use visualization to:
- Confirm detected issues: Visually verify that flagged images indeed contain annotation errors.
- Understand error types: Determine whether an issue is an overlooked object, a swapped label, or a badly located box.
- Guide corrections: Provide annotators with visual context for fixing detected issues.
- Generate reports: Save annotated images for documentation and quality review processes.
Theoretical Basis
The visualization approach is grounded in the following rendering strategy:
Ground truth rendering: Draw ground truth bounding boxes in a distinct color (e.g., green). Annotate each box with its class label.
Prediction rendering: Draw predicted bounding boxes in a contrasting color (e.g., red). Optionally filter predictions by a confidence threshold to reduce visual clutter from low-confidence detections.
Overlay composition: In overlay mode, render both sets of boxes on the same image canvas. Use transparency or distinct line styles to ensure both are visible when boxes overlap.
For each box in ground_truth:
draw_rectangle(image, box.coords, color=green)
draw_label(image, box.class_name, position=box.top_left)
For each box in predictions (where confidence > threshold):
draw_rectangle(image, box.coords, color=red)
draw_label(image, box.class_name, position=box.top_left)
Optional saving: The rendered image can be saved to disk for offline review or report generation.