Analyzing Your Metadata
Use the Analysis View to quickly summarize your dataset across multiple metadata dimensions.
Last updated
Use the Analysis View to quickly summarize your dataset across multiple metadata dimensions.
Last updated
The Analysis View, accessible by clicking on the bar chart icon () in the left hand navigation bar, allows you to understand the distribution of your dataset and your models' performance across frame and crop level metadata.
The Analysis View combines arbitrarily configured charts summarizing frame, label or inference metadata with cascading filters.
These capabilities are an entry point for several common workflows in Aquarium, including:
Identifying underrepresented segments of your dataset to queue for further curation with collection campaigns.
Identifying metadata-driven trends in model performance to track via regression tests.
Each card in the Analysis View is an individually configurable chart that can be used to plot any metadata at the frame, label or inference level.
The X axis summarizes categorical, numerical, or boolean metadata attached to any of Frames, Labels or Inferences.
The Y axis aggregates frame, label or inference counts, frame level model performance (F1, Precision, Recall), or numerical metadata using common aggregation types (count, average, maximum, minimum).
You can build as many charts as you'd like or edit any existing chart by clicking the edit icon next to the chart title.
Defining the X Axis
Choose the Metadata Level from one of frame, label or inference.
Choose the Metadata Field from the available options.
These are pre-filtered based on your metadata level selection.
Defining the Y Axis
Choose the Aggregate Type from one of count, average, maximum or minimum.
Note that count will by default count the number of elements at your chosen metadata level.
Choose the Aggregate Field from one of the available options.
Note that the aggregate field option will only appear if you select average, maximum or minimum. Otherwise it defaults to the count of elements.
Note that model performance statistics are only available if your Metadata Level is frame.
All configured charts support interactive, cascading filters.
To apply a filter, select any combination of bars within a chart.
Filters cascade downwards. For a given chart, any applied filter also applies to all charts below it.
With a frame level filter applied:
Subordinate frame charts will only plot frames in the filtered frame set.
Subordinate label charts will only plot labels within frames in the filtered frame set.
Subordinate inference charts will only plot inferences within frames in the filtered frame set.
With a label level filter applied:
Subordinate frame charts will plot frames with at least one label matching the filtered label condition.
Subordinate label charts will plot only labels within the filtered label set.
Subordinate inference charts will plot only inferences matched to the filtered labels at the configured IOU / confidence threshold.
With an inference level filter applied:
Subordinate frame charts will plot frames with at least one inference matching the filtered inference condition.
Subordinate label charts will plot only labels matched to the filtered inferences at the configured IOU / confidence threshold.
Subordinate inference charts will plot only inferences within the filtered inference set.
Clicking the explore button () on any chart allows you to interact with the filtered set of results anywhere else in the app.
