- 21 Dec 2022
- 2 Minutes to read
Use Case: False Positives of a Particular Class
- Updated on 21 Dec 2022
- 2 Minutes to read
There may be times when a Model has multiple False Positives for a particular Class. This article will help you:
- Identify the signal and error
- Diagnose the root cause
- Implement the fix
Identify the Signal and Error
When you analyze your Model and notice a high amount of False Positives, look through your data to confirm these False Positives.
If the Model has several False Positives, you may notice these signs:
- Several False Positives of a particular Class
- The Model identifies a specific Class, but that Class is not in the image
Diagnose the Root Cause
You can diagnose the root cause by parsing through your training data. For this use case of False Positives of a particular Class, there are five potential causes:
- There are too many of one type of example, like NG examples, so the dataset induces bias in the Model.
- The Segmentation output has small patches/pixels of "noise".
- The Confidence Threshold is too low.
- The environmental noise causes OK regions to look like detectable areas.
- The labeling convention does not differentiate between OK and NG regions, or the labelers accidentally drew extra labels.
Take a look at the example image below. The Model detected several areas that it should not have; thus, this image has multiple False Positives.
Now that you've learned how to identify the signal and diagnose the cause, you can iterate your data to achieve higher accuracy by following these tips:
|Too many NG examples
|Add more counter-examples, or apply Class weights to loss
|Segmentation output has small patches/pixels of "noise"
|Apply postprocessing to filter noise in the segmentation output. See sample noise filtering scripts here.
|Confidence Threshold is too low
|Increase the Confidence Threshold
|Environmental noise causes OK regions to look like defects
|Try removing the noise from the data collection process, or use data augmentation to simulate
|Labeling convention does not differentiate between OK and NG regions or the labelers accidentally drew extra labels