HehoFeature size is a setting in these modules: Anomaly DetectorDetectorSurface DetectionOCR
During annotations, you can change the feature size with a slider. A grid is going to appear showcasing the expected size of image features. Based on the grid you can determine what size of defects the model will be able to recognize.
Finding the optimal Feature size for your use case might require multiple trainings. Have a look at the following examples to help you determine how to optimally set Feature size.
It is possible to keep the Feature size at 32px at all times, which will work. However, you might be unnecessarily wasting time on training and evaluation.
Using high feature size:
Provides the model with less detail and context about the image.
| Tip |
|---|
Faster training time and better inference performance. |
| Warning |
|---|
Worse precision on small details and defects in the image. |
Notice the filled holes in the nuts. With higher feature size the shape of the objects is correctly found in a very short time. The drawback is lower shape precision.
...
Using low feature size:
Provides the model with more detail and context about the image.
| Tip |
|---|
Better precision on small details and defects in the image. |
| Warning |
|---|
Slower training time and slower inference performance. |
Notice that the precision of the shape is much better now. The drawback was longer training time and inference time. To further boost the shape precision a more detailed image would be needed.
...