Similarity search
  • 19 Dec 2022
  • 1 Minute to read
  • Dark
    Light
  • PDF

Similarity search

  • Dark
    Light
  • PDF

Article Summary

Similarity search allows driving a visual search or a model error characteristics-based search for images/video frames of interest in a large dataset. For visual search, there are two types of searches that are supported:
  1. Full image search: The full image is used as the search candidate to find matching images.
  2. Patch search: If the point of interest is a subset of the entire image, then patch search allows searching based on the specific point of interest. E.g., suppose the intent is to find all speed limit signs in an automotive dataset. In that case, a patch search allows expressing intent specifically on the speed limit board, and the search will find images with similar regions present anywhere in the image.

The below video provides an introduction to the similarity search process. 


Patch search is available only if the 'Patch' featurizer is specified when creating the dataset.

Was this article helpful?