Skip to content

How to run Find Similar

What can I do with Find Similar?

  • Group visually similar detections to reduce labeling effort.
  • Propagate known labels to unknown detections.
  • Prepare embeddings for semi-supervised LDA labeling.
  • Visually inspect and correct clusters via FiftyOne's UMAP brain view.

1. Confirm prerequisites

  • Run Unified Detection first; each detection must contain a leip_embedding.
Inspect detections

Inspect Unified Detections

2. Launch and Execute the Find Similar operator

  • Select the Find Similar operator in the Browse Operations menu and click Execute.
Execute Find Similar

Execute Find Similar

Default clustering behavior

The default (and recommended) clustering method combines k-means clustering and Linear Discriminant Analysis (LDA). k-means clustering performs unsupervised grouping of detection embeddings into clusters of visually similar objects; LDA trains a lightweight classifier to extend labels to unlabeled detections. To run k-means clustering only, select that option under Show Advanced Options.

3. Inspect embedding visualization

  • Open the embedding visualization by selecting brain_key='find_similar_clusters'.
Open embedding visualization

Open embedding visualization

  • Confirm clusters are coherent (similar objects grouped together).
Confirm cluster coherency

Confirm cluster coherency

  • Check unknowns: many should now have labels from majority voting or LDA.

Tip

After modifying a few labels manually, re-run Find Similar to allow LDA to unlock and propagate corrections automatically.

Troubleshooting

If you run into problems, consult the troubleshooting guide.