App takes images project (or dataset) as an input and produces an “Interactive co-occurrence matrix” that has the following dimensions: row_number = number of classes, col_number = number of classes. Each cell of the matrix shows how often a pair of classes (say person and car) appears together (how many images that simultaneously contain at least 1 person and at least 1 car). Each cell is clickable to open corresponding images.
Additional comments:
- This app is good for data exploration since it allows to see a big picture (co-occurrence statistics) as well as to navigate to the images of interest
- This App might be used to find “suspicious annotations”. If annotator has confused the class, we might see it as an “unexpected value” in the cell of “coexistence matrix”
- Gradient based colors of the matrix’ cells might be useful (will be added in next version)
Step 1: Add app to your team from Ecosystem if it is not there
Step 2: Run app from the context menu of images project or dataset: context menu -> Reports -> Classes Co-Occurrence Matrix
Step 3: Wait until the app is started, press Open button in Workspace tasks. You don't need to wait untill all images are processed: if open button is enabled then click it.
Step 4: Explore you data with interactive table: click on cells to access corresponding images and open them in labeling UI.
Step 5: App saves link to report to team files: /reports/classes-co-occurrence/<project id>_<project_name>.lnk. Link to generated report also available in task output column.
Step 6: Stop application once you finished with it. App can be stopped from tasks list or from application UI.
Example of the results:
| Stop from App UI | Stop from workspace tasks page |
|---|---|
 |
 |