Manage, Clean & Curate Visual Data - Fast and at Scale.

An unsupervised and free tool for image and video dataset analysis.
Explore the docs »
Features · Report Bug · Read Blog · Quickstart · Enterprise Edition · About us

🔥 We've released fastdup V1.0! View the release notes here.

What's Included

fastdup analyzes your image/video dataset for potential issues such as -

fastdup works on both labeled and unlabeled data. Additional features include -

Why fastdup?

• Quality: Find and remove anomalies and outliers from your dataset, including duplicates and similar images and videos at a large scale.
• Cost: Reduce data operation costs by intelligently sampling high-quality or novel datasets before labeling and assessing labeled data quality.
• Scale: fastdup's C++ graph engine is highly efficient and can handle up to 400M images on a single CPU machine.

Setting up

Prerequisites

Supported Python versions:

Supported operating systems:

Installation

Option 1 - Install fastdup via PyPI:

# upgrade pip to its latest version
pip install -U pip

# install fastdup
pip install fastdup
    
# Alternatively, use explicit python version (XX)
python3.XX -m pip install fastdup 

Option 2 - Install fastdup via an Ubuntu 20.04 Docker image on DockerHub:

docker pull karpadoni/fastdup-ubuntu-20.04

Detailed installation instructions and common errors here.

Getting Started

Run fastdup with only 3 lines of code.

Visualize the result.

Here are 8 lines of code you'll need in most cases.

import fastdup

fd = fastdup.create(work_dir, images_dir)
fd.run()

fd.vis.duplicates_gallery()    # create a visual gallery of found duplicates
fd.vis.outliers_gallery()      # create a visual gallery of anomalies
fd.vis.component_gallery()     # create a visualization of connected components
fd.vis.stats_gallery()         # create a visualization of images statistics (for example blur)
fd.vis.similarity_gallery()    # create a gallery of similar images

View the API docs here.

Learn from Examples

	Quick Dataset Analysis: In this example, learn how to quickly analyze a dataset for potential issues. Identify duplicates, outliers, dark/bright/blurry images, and cluster similar images with only a few lines of code. If you're new, start here.
	DINOv2 Embeddings: In this example, learn how to use DINOv2 models to visualize image embeddings of your dataset. Runs on CPU!
	Cleaning Image Dataset: In this tutorial, learn how to clean a dataset from broken images, duplicates, outliers, and identify dark/bright/blurry images.
	Analyzing Labeled Image Classification Dataset: In this tutorial, learn how to analyze a labeled image classification dataset for potential issues. We use the Imagenette dataset, a 10-class, 13k image subset of ImageNet as a working example.
	Analyzing Labeled Object Detection Dataset: In this tutorial learn how to load and analyze an object detection dataset with labeled bounding boxes and classes. We use the mini-coco dataset as a working example. Learn how to discover duplicates, outliers, and possible mislabeled bounding boxes.

Advanced Features

The following are advanced functionalities of fastdup which are still in the beta testing phase. Sign up for free to be a beta tester and get early access. Drop us an email at info@visual-layer.com .

	Face Detection Video Analysis: In this tutorial, learn how to use fastdup with a face detection model to detect and crop from videos. Following that we analyze the cropped faces for issues such as duplicates, near-duplicates, outliers, bright/dark/blurry faces.
	YOLOv5 Object Detection Video Analysis: In this tutorial, learn how to use fastdup with a pre-trained yolov5 object detection model to detect and crop from videos. Following that we analyze the cropped objects for issues such as duplicates, near-duplicates, outliers, bright/dark/blurry objects.
	Satellite Image Analysis: In this tutorial, learn how to use fastdup to load 16-bit grayscale satellite image, work with rotated bounding boxes, understand your dataset, find issues with the data and check the quality of annotations.
	Surveillance Camera Analysis: In this tutorial, learn how to use fastdup to analyze surveillance camera videos, caption the activity inside the videos and detect indoor/ outdoor.
	Image Search: In this tutorial, learn how to use fastdup to search through large image datasets for duplicates/similar images using a query image. Runs on CPU!

Getting Help

Get help from the fastdup team or community members via the following channels -

• Slack.
• GitHub issues.
• Discussion forum.

Community Contributions

The following are community-contributed blog posts about fastdup -

• Master Data Integrity to Clean Your Computer Vision Datasets.
• fastdup: A Powerful Tool to Manage, Clean & Curate Visual Data at Scale on Your CPU - For Free.
• Clean Up Your Digital Life: How I Found 1929 Fully Identical Images, Dark, Bright and Blurry Shots in Minutes, For Free.
• The weighty significance of data cleanliness — or as I like to call it, “cleanliness is next to model-ness” — cannot be overstated.

What our users say

License

fastdup is licensed under Creative Commons 4.0 license. See LICENSE.

For any queries, reach us at info@visual-layer.com

Disclaimer

Usage Tracking

We have added an experimental crash report collection, using sentry.io. It does not collect user data other than anonymized IP address data, and it only logs fastdup library's own actions. We do NOT collect folder names, user names, image names, image content only aggregate performance statistics like total number of images, average runtime per image, total free memory, total free disk space, number of cores, etc. Collecting fastdup crashes will help us improve stability.

The code for the data collection is found here. On MAC we use Google crashpad.

It is always possible to opt out of the experimental crash report collection via either of the following two options:

• Define an environment variable called SENTRY_OPT_OUT
• or run() with turi_param='run_sentry=0'

About Visual-Layer

fastdup is founded by the authors of XGBoost, Apache TVM & Turi Create - Danny Bickson, Carlos Guestrin and Amir Alush.

Learn more about Visual Layer here.