Understanding Clusters

Sifting through large collections of images and videos is a time consuming task. To address this challenge, Visual Layer groups similar images or Core Frames from videos into Clusters for you.

Clusters are a tool designed to help you quickly explore, filter, tag and sort through vast amounts of visual data. Images and Core Frames are grouped together according to visual similarity, as are Objects with similar visual properties.

Cluster previews

The Cluster preview provides a sneak peek at the data within prior to its exploration:

• Number of items - Total number of items that comply with the current filtering criteria
• Prominent labels - Up to 3 of the most prominent Image and Object labels
• Representing Images - Up to 16 representing images and Core Frames

This is how Cluster previews look for Images and/or Video Frames:

This is how Cluster previews look for Objects (Objects are cropped):

A click on a Cluster preview card opens the Cluster Details page, which shows more information about the data inside it than the preview.

Cluster details - Images View

When looking at the details of a Cluster of Images and/or Video Frames, you can see:

• A preview card for every Image or Video Frame in the cluster which includes:
  • The file name
  • The assigned label/s
  • A preview Image
  • The name of the Video file (for Video Frames only)
• A metadata summary (right pane), which includes:
  • Quality issues pertaining to this specific cluster
  • Top User Tags assigned to items in this Cluster
  • Top Image and Object labels assigned to items in this Cluster

Any filters applied prior to inspecting a cluster will be applied to the Cluster Details Page.

Cluster details - Objects

When looking at the details of a Cluster of Objects, you can see:

• A preview card for every Object in the cluster which includes:
  • The file name of the Video or Image from which this Object is derived
  • The Object label
  • A preview image - With the bounding box of the Object highlighted over the image
• A metadata summary (right pane), which includes:
  • Quality issues pertaining to this specific cluster
  • Top User Tags assigned to items in this Cluster
  • Top Image and Object labels assigned to items in this Cluster

Like with Images, any filters applied prior to inspecting a cluster will be applied to the Cluster Details Page.

Clustering granularity

The level of similarity that determines the clustering is adjustable - the Granularity knob at the bottom right side of the screen can be used to alter it.

For more semantic and loose clustering (lower similarity threshold), move it to the left (this is the default granularity):

If you’d like clustering to only be applied to very similar images, move the knob to the right: