Threshold values inform the classification of the images in the dataset by providing scores and confidence levels.
After the training phase, scores are calculated for every training sample and presented as a histogram; good samples are marked with green bars and bad samples with red bars. All images with scores between 0-T1 are marked Good, and images with a score above T2 are classified as Bad. If the score is between T1 and T2, the result is [Confident: No].
Training with many samples from both groups is recommended to achieve a more robust threshold.
The histogram tool displays green bars representing correct samples and red bars representing anomalous samples. T (the midpoint of T1 and T2) marks the main threshold, and T1 and T2 define the area of uncertainty [Confident: No].
Images that have scores within 0-T1 are marked as Good, [Confident: Yes]
Images that have scores within T1-T are marked as Good, [Confident: No]
Images that have scores within T-T2 are marked as Bad, [Confident: No]
Images that have scores within T-T2 are marked as Bad, [Confident: Yes]
If both thresholds are equal (T1=T2=T), there is no area of uncertainty. Results are marked as Confident: Yes
The image viewer in Aurora Focus provides information on Tool Time, Confident, and Score. [Confident: No] indicates the score is close to the threshold. In this case, perform another inspection.
The following histogram displays well-separated groups, indicating that the model has good accuracy:
Uncertainty Threshold
The following histogram displays groups in close proximity, indicating that the model has poor accuracy:
Good and Bad Model Accuracy Examples
The following histogram provides a real-world example in Aurora Deep Learning Editor:
Aurora Deep Learning Editor Example
This example shows 36 images in the model with clear separation between Good and Bad images. The range of scores for Good images is 7.04-31.12, and the range of scores for Bad images is 56.20-106.80. The yellow separation between Good and Bad images indicates that the model has good accuracy.
