Find Duplicates in FiftyOne Datasets

Overview

Find and remove duplicate or near-duplicate images using FiftyOne's brain similarity operators. Uses deep learning embeddings to identify visually similar images.

Use this skill when:

• Removing duplicate images from datasets
• Finding near-duplicate images (similar but not identical)
• Clustering visually similar images
• Cleaning datasets before training

Prerequisites

• FiftyOne MCP server installed and running
• @voxel51/brain plugin installed and enabled
• Dataset with image samples loaded in FiftyOne

Key Directives

ALWAYS follow these rules:

1. Set context first

set_context(dataset_name="my-dataset")

2. Launch FiftyOne App

Brain operators are delegated and require the app:

launch_app()

Wait 5-10 seconds for initialization.

3. Discover operators dynamically

# List all brain operators
list_operators(builtin_only=False)

# Get schema for specific operator
get_operator_schema(operator_uri="@voxel51/brain/compute_similarity")

4. Compute embeddings before finding duplicates

execute_operator(
    operator_uri="@voxel51/brain/compute_similarity",
    params={"brain_key": "img_sim", "model": "mobilenet-v2-imagenet-torch"}
)

5. Close app when done

close_app()

Complete Workflow

Step 1: Setup

# Set context
set_context(dataset_name="my-dataset")

# Launch app (required for brain operators)
launch_app()

Step 2: Verify Brain Plugin

# Check if brain plugin is available
list_plugins(enabled=True)

# If not installed:
download_plugin(
    url_or_repo="voxel51/fiftyone-plugins",
    plugin_names=["@voxel51/brain"]
)
enable_plugin(plugin_name="@voxel51/brain")

Step 3: Discover Brain Operators

# List all available operators
list_operators(builtin_only=False)

# Get schema for compute_similarity
get_operator_schema(operator_uri="@voxel51/brain/compute_similarity")

# Get schema for find_duplicates
get_operator_schema(operator_uri="@voxel51/brain/find_duplicates")

Step 4: Compute Similarity

# Execute operator to compute embeddings
execute_operator(
    operator_uri="@voxel51/brain/compute_similarity",
    params={
        "brain_key": "img_duplicates",
        "model": "mobilenet-v2-imagenet-torch"
    }
)

Step 5: Find Near Duplicates

execute_operator(
    operator_uri="@voxel51/brain/find_near_duplicates",
    params={
        "similarity_index": "img_duplicates",
        "threshold": 0.3
    }
)

Threshold guidelines (distance-based, lower = more similar):

• 0.1 = Very similar (near-exact duplicates)
• 0.3 = Near duplicates (recommended default)
• 0.5 = Similar images
• 0.7 = Loosely similar

This operator creates two saved views automatically:

• near duplicates: all samples that are near duplicates
• representatives of near duplicates: one representative from each group

Step 6: View Duplicates in App

After finding duplicates, use set_view to display them in the FiftyOne App:

Option A: Filter by neardupid field

# Show all samples that have a near_dup_id (all duplicates)
set_view(exists=["near_dup_id"])

Option B: Show specific duplicate group

# Show samples with a specific duplicate group ID
set_view(filters={"near_dup_id": 1})

Option C: Load saved view (if available)

# Load the automatically created saved view
set_view(view_name="near duplicates")

Option D: Clear filter to show all samples

clear_view()

The findnearduplicates operator adds a neardupid field to samples. Samples with the same ID are duplicates of each other.

Step 7: Delete Duplicates

Option A: Use deduplicate operator (keeps one representative per group)

execute_operator(
    operator_uri="@voxel51/brain/deduplicate_near_duplicates",
    params={}
)

Option B: Manual deletion from App UI

1. Use setview(exists=["neardup_id"]) to show duplicates
2. Review samples in the App at http://localhost:5151/
3. Select samples to delete
4. Use the delete action in the App

Step 8: Clean Up

close_app()

Available Tools

Session View Tools

Brain Operators for Duplicates

Use listoperators() to discover and getoperator_schema() to see parameters:

Common Use Cases

Use Case 1: Remove Exact Duplicates

For accidentally duplicated files (identical bytes):

set_context(dataset_name="my-dataset")
launch_app()

execute_operator(
    operator_uri="@voxel51/brain/find_exact_duplicates",
    params={}
)

execute_operator(
    operator_uri="@voxel51/brain/deduplicate_exact_duplicates",
    params={}
)

close_app()

Use Case 2: Find and Review Near Duplicates

For visually similar but not identical images:

set_context(dataset_name="my-dataset")
launch_app()

# Compute embeddings
execute_operator(
    operator_uri="@voxel51/brain/compute_similarity",
    params={"brain_key": "near_dups", "model": "mobilenet-v2-imagenet-torch"}
)

# Find duplicates
execute_operator(
    operator_uri="@voxel51/brain/find_near_duplicates",
    params={"similarity_index": "near_dups", "threshold": 0.3}
)

# View duplicates in the App
set_view(exists=["near_dup_id"])

# After review, deduplicate
execute_operator(
    operator_uri="@voxel51/brain/deduplicate_near_duplicates",
    params={}
)

# Clear view and close
clear_view()
close_app()

Use Case 3: Sort by Similarity

Find images similar to a specific sample:

set_context(dataset_name="my-dataset")
launch_app()

execute_operator(
    operator_uri="@voxel51/brain/compute_similarity",
    params={"brain_key": "search"}
)

execute_operator(
    operator_uri="@voxel51/brain/sort_by_similarity",
    params={
        "brain_key": "search",
        "query_id": "sample_id_here",
        "k": 20
    }
)

close_app()

Troubleshooting

Error: "No executor available"

• Cause: Delegated operators require the App executor for UI triggers
• Solution: Direct user to App UI to view results and complete deletion manually
• Affected operators: findnearduplicates, deduplicatenearduplicates

Error: "Brain key not found"

• Cause: Embeddings not computed
• Solution: Run computesimilarity first with a brainkey

Error: "Operator not found"

• Cause: Brain plugin not installed
• Solution: Install with downloadplugin() and enableplugin()

Error: "Missing dependency" (e.g., torch, tensorflow)

• The MCP server detects missing dependencies automatically
• Response includes missingpackage and installcommand
• Example response:

``json { "errortype": "missingdependency", "missingpackage": "torch", "installcommand": "pip install torch" } ``

• Offer to run the install command for the user
• After installation, restart MCP server and retry

Similarity computation is slow

• Use faster model: mobilenet-v2-imagenet-torch
• Use GPU if available
• Process large datasets in batches

Best Practices

1. Discover dynamically - Use listoperators() and getoperator_schema() to get current operator names and parameters
2. Start with default threshold (0.3) and adjust as needed
3. Review before deleting - Direct user to App to inspect duplicates
4. Store embeddings - Reuse for multiple operations via brain_key
5. Handle executor errors gracefully - Guide user to App UI when needed

Performance Notes

Embedding computation time:

• 1,000 images: ~1-2 minutes
• 10,000 images: ~10-15 minutes
• 100,000 images: ~1-2 hours

Memory requirements:

• ~2KB per image for embeddings
• ~4-8KB per image for similarity index

Resources

• FiftyOne Brain Documentation
• Brain Plugin Source

License

Copyright 2017-2025, Voxel51, Inc. Apache 2.0 License