The similarity search is run against all points in a job by default. This can be changed by running a search on a filtered search space based on the conditions available in the following drop-down. The following image shows the filtering conditions available for selection.
The selection in the search tab is whether to apply one or more filters, and the exact specification of these filters is done in the selection tab, as shown below. Please find details on specifying the filters in Select and Refine.
Use cases
- Tunable: In a video dataset, a tunable with 'uniform' sampling and a sampling fraction of N can be set as below to reduce search space by removing similar-looking consecutive frames. In the search panel, if 'tunable' is selected from the drop-down, then the search will only be restricted to the sampled frames.
- Catalog refine: You can restrict search based on a specific attribute in the catalog(say the class name) by setting up the catalog filter and selecting 'catalog refine' in the search screen drop-down.
- Selection panel filters: The plot view is expected to show the distinct separation of visual characteristics among the different clusters. Let us say there are 2 clusters formed with day-time and night-time images. A filter can be specified to choose a particular cluster, and a subsequent search can be set to apply this filter.
