Dataset Ninja: Search and Explore Computer Vision Datasets

Dataset Ninja is a free service for searching and exploring computer vision datasets. Every dataset has a preview gallery, interactive statistics such as class balance and a confusion matrix, and one common annotation format. It was created by Supervisely's co-founders, and you can send any dataset to Supervisely to train models.

Hunting for datasets and preparing them takes time. Dataset Ninja, at DatasetNinja.com, is meant to make that step short. It is free for the entire Machine Learning community, and the co-founders of Supervisely built it.

Why is finding computer vision datasets hard?

Hunting down training datasets can be really tough and frustrating. There are many dataset resources like:

1. Kaggle

2. Zenodo

3. Mendeley Data

4. Hundreds of awesome lists and GitHub repositories like this.

Your ML journey begins with searching the web and unstructured dataset catalogs, downloading huge archives, deciphering intricate annotation formats and licenses. Next, you'll dive into Python scripting to create visualizations and convert the data into a more appropriate annotation format. Also, remember to use Jupyter notebooks to uncover the complex statistics concealed within the data.

Dataset page has every important piece of information about it

These challenges are just a fraction of the broader tasks you'll encounter. Being a proficient data scientist requires a set of tools to thoroughly comprehend all aspects of your training data.

What is Dataset Ninja and what can you do on it?

DatasetNinja.com provides convenient interactive analytical tools and visualizations for dataset experts and data scientists out-of-the-box. Watch this 5 minutes introduction video that illustrates the main concepts of Dataset Ninja:

Dataset Ninja is built on the following principles:

How can you explore dataset statistics interactively? 📊

Besides quickly previewing all images with annotations in a thumbnail gallery, you can also explore class balance, confusion matrix, spatial distribution of labels and other aggregated statistics. Just click the matrix cell or table row to open all corresponding images. These interactivity opens up a new way to filter and explore datasets.

There are countless interactive statistics and charts to explore

Which annotation format do the datasets use? 📦

All datasets now have converters to single annotation JSON format — Supervisely. It comes with easy-to-use Python SDK - just pip install supervisely and has Developer portal with nice documentation and examples.

How do you search for a dataset? 🔎

Search by industry, class name or by license - now you do not have to go digging around hundreds of unstructured lists and websites. Dataset page gives all relevant information, links, summary and license references. Everything you need is there.

Catalog has tons of filters and sort options

How do you train on a Dataset Ninja dataset in Supervisely? 🤝

Native integration with Computer Vision platform Supervisely allows you to make the way from labeled data to model training in just a few clicks. Once a dataset is in your account, you can explore and analyze it and import or convert data with the import docs. For example, you can train state-of-the-art models like YOLOv8 or use any of 250+ tools from Supervisely Ecosystem in your machine learning research.

Just one click — and you can train any dataset in Supervisely

How often are new datasets added? 🔥

Dedicated team of python engineers and dataset experts on a daily basis collects and systemizes CV datasets, implements data converters and calculates all statistics and visualizations. Any contributions from Computer Vision Community like ideas, dataset suggestions or Python SDK improvements are welcomed. We appreciate any help and hope that collaborative work of the entire community can make a huge impact on the AI field by making data structured and easily accessible for everyone.

Dataset Ninja vs manual dataset hunting: when to use which?

The table compares the manual route described above with a Dataset Ninja dataset page.

	Manual search (Kaggle, Zenodo, Mendeley Data, GitHub lists)	Dataset Ninja
Finding a dataset	Unstructured catalogs and web search	Search by industry, class name or license
Preview	Often a download of the archive first	Thumbnail gallery with annotations
Statistics	Your own Python scripts and notebooks	Class balance, confusion matrix and spatial distribution, interactive
Annotation format	Differs per dataset	One Supervisely JSON format
Training	Convert, then set up your own pipeline	Add to Supervisely in one click

Use Dataset Ninja to find and inspect a dataset quickly. Go to the original source when you need a dataset that the catalog does not list yet.

Conclusion

The main goal of Dataset Ninja is to centralize and structure the world of Computer Vision datasets in one place, build the central hub where datasets reside and provide the best tools for dataset exploration and analysis.

It is a well-known fact that the quality of training data significantly influences model accuracy. We firmly believe that interactive visualizations and comprehensive statistics will empower data scientists to gain a deeper understanding of their training data, enabling efficient quality assurance and ultimately leading to the training of more accurate neural networks.