Models
This model crops the image to Zone 5 before any defect detection occurs.
What is Bottom Foreground?
This is the dataset split can be used to adjust settings like learning rate and check for overfitting during training.
What is the validation set?
This metric measures how often the model is correct when it flags something as a defect.
What is Precision?
This is where trained model versions are moved to and from Gadgets.
What is the Model Management page in GadgetApps?
This is how many cameras and lights are present on the machine, in total.
What are 4 cameras and 4 lights?
This model is trained exclusively on good bottles and flags anything that deviates from them.
What is Bottom AD (Anomaly Detection)?
This training parameter, if set too low, may cause training to stop before the model has had enough time to learn meaningful patterns.
What is Patience?
A high score on this metric means the model is finding objects in roughly the right place, but says nothing about how precisely the boundaries are drawn.
What is AP50?
This is the alarm on the machine for when the maximum number of rejects for a particular defect was reached.
What is Defect Limit Alarm?
This is the specific section within Label Studio Settings where an annotation project can be permanently deleted.
What is the Danger Zone?
Unlike other defect models, this model's annotations are used specifically to reduce false positives rather than to identify defects.
What is Neg Mask Bottom?
This is what it means when both training loss and validation loss remain high and flat throughout training.
What is underfitting?
This visualization makes it easy to see which defect classes the model is mixing up with one another.
What is the Confusion Matrix?
These are the three conveyors on the machine.
What are Bi-Flo, Gap Transport, and Outfeed?
When annotating good bottles for the Bottom AD model, these annotations are used to perform this specific technique to reduce false positives.
What is negative masking?
The Bottom Foreground and Side Foreground models both use this algorithm.
What is YOLO11 Detect?
This training issue happens when the model memorizes instead of generalizes due to insufficient variability.
What is overfitting?
These are the different loss curves that need to be analyzed to evaluate the model version that was trained.
What are training loss and validation loss curves?
This is how anomaly detections appear differently from all other defects on the HMI Defect Tracking page.
What is they are recorded as "Other"?
What are the three preprocessing settings that can be seen in the Bottom Defect model.
What is Rotate, Crop to Label, and Resize?
This is the algorithm used by both anomaly detection models in the pipeline.
What is PatchCore?
When training stops early despite loss still improving, this parameter is the cause.
What is Patience set too low?
This metric evaluates overlap between predicted and actual annotattions.
What is IoU?
What is the PLC?
When no dedicated test set is available during model setup, the documentation states this can be used in its place.
What is the training data?