Image similarity search tutorial

To use the similarity search tool, you will need to first choose from either 1 versus N or N versus N.

1 v N finds all similar images in your dataset to your query image

N v N finds all similar image pairs in your dataset

The ‘number of results’ box allows you to choose how many results you see. If you enter 0, the results will be unlimited, with all images in the dataset sorted by similarity

Users can enter a minimum score threshold (%) , so that only images with similarity prediction above this threshold will be displayed.

If you are using the 1 vs N model, you will need to start up by uploading your query image. Then, click on start, and this will run a similarity search against all images in the dataset. The results will be filtered by top scoring images by similarity.

You can download these results in CSV format, JSON format, or download the top scoring images by similarity.

On the left hand side, you will be able to see the image similarity search results history. This search history can be easily accessed at any time, by clicking on Image Similarity Search at the top, and then click ‘Show History’.

Rather than uploading a specific query image, you can perform a similarity search on any image already in your dataset. To do this, simply right click on the image, and then select similarity search.

To find the most similar image pairs in your dataset, run the N vs N model. Enter the number of results and score threshold as before. Then, click Start.

In N vs N, there is no query image. Instead, the result images are split into pairs, and ranked on their similarity score. This can be useful to remove any duplicates in a dataset