SiaSearch v0.8: Introducing Similarity Search
In SiaSearch v0.8, we’ve added a feature that allows users to search for frames and objects of interest based on visual similarity.
🔎 Similarity Search
When curating a training dataset, have you ever wanted to find similar examples to a specific frame or object? To give you an idea, imagine you’ve identified a failure case for your model, and now want to search for similar instances in order to diagnose the issue.
We are happy to announce that in SiaSearch v0.8, users can now search for visually similar crops over specified regions of images.
Similarity search allows users to go beyond the constraints of metadata search to:
- Run qualitative search based on visual inspection
- Find items not covered by annotations / model predictions
- Curate datasets faster once an ideal sample is obtained
In order to search for similar crops, users can simply navigate to the Frames view in the search results. Once you’ve found a frame of interest, select the similarity search option and draw the region directly onto the image. By running similarity search, you’ll be able to obtain visually similar crops. After the results have been obtained, you can add items of interest to a collection to access them at a later time.
Please note that currently, entire frames will be added to the collection. Support for frame crops will be added soon.
Under the hood, SiaSearch makes use of a rich feature store on an intelligent sampling of image crops to power similarity search. With this powerful functionality, you can easily query and sort datasets to find similar examples.
The SiaSearch team is constantly working on new features to improve the SiaSearch product. To stay up to date on SiaSearch releases as they are shipped, sign up to receive our newsletter.
