Find Mislabeled Data Using Visual Layer Auto-Detection
Understand why data was detected as "mislabeled" by Visual Layer
Visual Layer analyzes the data in order to detect and surface suspected mislabeled data. This process involves identifying discrepancies between the Visual Layer analysis and the provided labeled.
Both Images and objects can be detected as mislabeled. Unlike image mislabels, object mislabels can result from incorrectly positioned or sized bounding boxes, leading to the capture of partial, occluded or multiple objects.
Visual Layer provides an explanation for the mis-detection of objects in the following form:
- Object likely mislabeled as cup, possible correction "bowl".
- Go to Dataset Inventoryview and select a dataset by clicking on it
- Select the required view: Images view or Objects view
- In the top filter bar, go to "Issues" filter, select "mislabeled" and apply the filter. Alternatively, use the right pane "Issues" section and select "mislabeled".
- The logical operator between items listed in the dropdown list of filters and within a section in the right pane is "OR".
- The logical operator between the filter types (User tags, Labels, Issues, etc.) is "AND", as well as for items selected in multiple separate sections in the right pane.
- The returned results comprise clusters of images or objects, depending on the selected view, each containing mislabeled item.
- Explore the cluster view to find the objects that were detected as mislabeled.
Visual Layer Explainability
- Navigate to the Image screen by clicking on an image. Identify any mislabels and review the explanation under the "Issues" section. Take note of the confidence level provided by Visual Layer.
Updated 4 months ago
Whatβs Next
To select data to send for relabeling using Visual Layer continue to the next page.