How to analyse FN / FP on detection annotations

What are FN / FP?

In the field of computer vision, FN (False Negative) and FP (False Positive) are two important concepts used to evaluate the performance of a model.

• FN（False Negative）：the model's failure to correctly identify the presence of an object or feature in an image.
• FP（False Positive）：the model's error of identifying non-existing objects or features as present in an image.

During the training process of deep learning models, both FN and FP have a significant impact on the model's performance, making them important indicators for model evaluation and optimization.

Prerequisites

1. Importing Ground Truth results into the dataset.
2. Importing user's predicted results set into the dataset.
3. Currently, only the Detection annotation type is supported for FN / FP analysis.

Usage

Enter analysis mode

1. Click on the 'Analysis' button in the top menu bar.
2. "Select the predicted results set that needs to be compared, and click on the 'Analyze FN / FP' button to enter analysis mode.

Analysis mode enter

Analyse

The system will pre-calculate the confidence threshold and corresponding FN / FP results under 0.1~0.9 Precision conditions.

Confidence compute

How to understand the boxes with different colored masks.

• Red masked box: represents FN, only appears in Ground Truth, indicating missed detection.
• Blue masked box: represents FP, only appears in the predicted results set, indicating false detection.
• Unmasked box: indicates that the predicted results set and Ground Truth match as expected.

Color sense

Sorting options

• Support selecting different preset confidence threshold options as the basis for sorting.
• Support sorting images from largest to smallest based on the number of FN or FP calculated from each image.

Sort options

Display options

• Support two comparison modes: overlay comparison and tiling comparison.
• Support controlling whether to display the boxes that match the predicted results set in Ground Truth, as well as the boxes that represent missed detections in the predicted results set.
• Support controlling whether to display the boxes that match Ground Truth in the predicted results set, as well as the boxes that represent false detections in the predicted results set.
• "Support dynamically displaying the corresponding calculation results under different confidence threshold options for the current sorting condition.

Display options

More demonstrations are shown below: