Exploring image similarity

Mine edge cases, duplicates, and check the quality of your labels with visual similarity search

Frequently, when distinctive characteristics arise within a dataset, identifying similar images becomes crucial (e.g., for relabeling or removal). Detecting these instances assists in assessing the thoroughness of data representation and the accuracy of labels, particularly in situations where certain classes may be underrepresented or labels could be inaccurately assigned. As datasets expand, manual identification of such cases becomes progressively challenging.

Leverage Encord Active's similarity search feature to effortlessly locate semantically akin images in your dataset. Upon identifying an edge case or duplicate, applying tags and executing actions such as relabeling or deletion can be performed.

Steps

  1. Access the Explorer page and locate the image or label of interest.
  2. Click the Similar items button associated with the selected sample. Encord Active will arrange the samples on the Explorer page, showcasing the most semantically similar images first.
    Displaying similar images based on the similarity search query

👍

Tip

To cancel the similarity search, you can click the X button located in the top right corner of the chosen sample or the RESET FILTERS button positioned near the natural language search bar.