Image Deduplication with FiftyOne¶

This recipe demonstrates a simple use case of using FiftyOne to detect and remove duplicate images from your dataset.

Setup¶

If you haven’t already, install FiftyOne:

[ ]:

!pip install fiftyone

This notebook also requires the tensorflow package:

[ ]:

!pip install tensorflow

Download the data¶

First we download the dataset to disk. The dataset is a 1000 sample subset of CIFAR-100, a dataset of 32x32 pixel images with one of 100 different classification labels such as apple, bicycle, porcupine, etc. You can use this helper script.

[1]:

from image_deduplication_helpers import download_dataset

download_dataset()

Downloading dataset of 1000 samples to:
        /tmp/fiftyone/cifar100_with_duplicates
and corrupting the data (5% duplicates)
Download successful

The above script uses tensorflow.keras.datasets to download the dataset, so you must have TensorFlow installed.

The dataset is organized on disk as follows:

/tmp/fiftyone/
└── cifar100_with_duplicates/
    ├── <classA>/
    │   ├── <image1>.jpg
    │   ├── <image2>.jpg
    │   └── ...
    ├── <classB>/
    │   ├── <image1>.jpg
    │   ├── <image2>.jpg
    │   └── ...
    └── ...

As we will soon come to discover, some of these samples are duplicates and we have no clue which they are!

Create a dataset¶

Let’s start by importing the FiftyOne library:

[2]:

import fiftyone as fo

Let’s use a utililty method provided by FiftyOne to load the image classification dataset from disk:

[3]:

import os

dataset_name = "cifar100_with_duplicates"
dataset_dir = os.path.join("/tmp/fiftyone", dataset_name)

dataset = fo.Dataset.from_dir(
    dataset_dir,
    fo.types.ImageClassificationDirectoryTree,
    name=dataset_name
)

 100% |████████████| 1000/1000 [1.2s elapsed, 0s remaining, 718.5 samples/s]

Explore the dataset¶

We can poke around in the dataset:

[4]:

# Print summary information about the dataset
print(dataset)

Name:           cifar100_with_duplicates
Media type:     image
Num samples:    1000
Persistent:     False
Info:           {'classes': ['apple', 'aquarium_fish', 'baby', ...]}
Tags:           []
Sample fields:
    filepath:     fiftyone.core.fields.StringField
    tags:         fiftyone.core.fields.ListField(fiftyone.core.fields.StringField)
    metadata:     fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.metadata.Metadata)
    ground_truth: fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Classification)

[5]:

# Print a sample
print(dataset.first())

<Sample: {
    'id': '5ff8dc665b5b9368e094de5a',
    'media_type': 'image',
    'filepath': '/tmp/fiftyone/cifar100_with_duplicates/apple/113.jpg',
    'tags': BaseList([]),
    'metadata': None,
    'ground_truth': <Classification: {
        'id': '5ff8dc665b5b9368e094de59',
        'label': 'apple',
        'confidence': None,
        'logits': None,
    }>,
}>

Create a view that contains only samples whose ground truth label is mountain:

[6]:

# Used to write view expressions that involve sample fields
from fiftyone import ViewField as F

view = dataset.match(F("ground_truth.label") == "mountain")

# Print summary information about the view
print(view)

Dataset:        cifar100_with_duplicates
Media type:     image
Num samples:    8
Tags:           []
Sample fields:
    filepath:     fiftyone.core.fields.StringField
    tags:         fiftyone.core.fields.ListField(fiftyone.core.fields.StringField)
    metadata:     fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.metadata.Metadata)
    ground_truth: fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Classification)
View stages:
    1. Match(filter={'$expr': {'$eq': [...]}})

[7]:

# Print the first sample in the view
print(view.first())

<SampleView: {
    'id': '5ff8dc675b5b9368e094e436',
    'media_type': 'image',
    'filepath': '/tmp/fiftyone/cifar100_with_duplicates/mountain/0.jpg',
    'tags': BaseList([]),
    'metadata': None,
    'ground_truth': <Classification: {
        'id': '5ff8dc675b5b9368e094e435',
        'label': 'mountain',
        'confidence': None,
        'logits': None,
    }>,
}>

Create a view with samples sorted by their ground truth labels in reverse alphabetical order:

[8]:

view = dataset.sort_by("ground_truth", reverse=True)

# Print summary information about the view
print(view)

Dataset:        cifar100_with_duplicates
Media type:     image
Num samples:    1000
Tags:           []
Sample fields:
    filepath:     fiftyone.core.fields.StringField
    tags:         fiftyone.core.fields.ListField(fiftyone.core.fields.StringField)
    metadata:     fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.metadata.Metadata)
    ground_truth: fiftyone.core.fields.EmbeddedDocumentField(fiftyone.core.labels.Classification)
View stages:
    1. SortBy(field_or_expr='ground_truth', reverse=True)

[9]:

# Print the first sample in the view
print(view.first())

<SampleView: {
    'id': '5ff8dc685b5b9368e094ea0f',
    'media_type': 'image',
    'filepath': '/tmp/fiftyone/cifar100_with_duplicates/worm/905.jpg',
    'tags': BaseList([]),
    'metadata': None,
    'ground_truth': <Classification: {
        'id': '5ff8dc685b5b9368e094ea0e',
        'label': 'worm',
        'confidence': None,
        'logits': None,
    }>,
}>

Visualize the dataset¶

Start browsing the dataset:

[10]:

session = fo.launch_app(dataset)

Narrow your scope to 10 random samples:

[11]:

session.view = dataset.take(10)