Exploring image similarity

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

Often, when you find characteristics of interest in your dataset you may want to identify other similar images (e.g., to send for relabelling or deletion). Identifying these cases can help you understand where your data is, e.g., underrepresented or mislabelled. As your dataset grows, finding such cases manually becomes increasingly difficult. Using Encord Active's similarity search you can easily find semantically similar images in your dataset. When have identified an edge case or a duplicate, you can tag it and export or delete it in the Actions tab.

Prerequisites: Dataset & Embeddings

Setup

If you haven't installed Encord Active, visit installation. In this workflow we will be using the COCO validation dataset.

Image similarity

Steps

1. Navigate to explorer tab

Navigate to the Data Quality > Explorer tab and select a quality metric in the top left menu to order your data by.

2. Select quality metric

Select a quality metric to index your images by:

3. Find similar images

Select an image of interest and click Show Similar Images. Encord Active shows the most semantically similar images below the image in a new tab.