Shutterstock’s image datasets are helping companies advance their machine learning tools, improve efficiencies and better serve their customers. Here’s how.
Computer vision — or training computers to identify objects in images as well as the human eye — is a field that gets more exciting each year with no signs of stopping. According to a recent report, the computer vision market will reach $19.1 Billion and grow at a rate of 7.6% per year by 2027.
Advancements in computer vision, alongside machine learning, are vastly changing the way companies operate, make decisions, serve their customers and improve our world.
For example, with computer vision technology:
- People can easily search images on their smartphone library by entering a keyword like “cat” or “ice skating” to find all relevant photos.
- Self-driving cars can operate safely by understanding their specific surroundings — including other cars, people, roads, and stop signs.
- Social media companies can rapidly identify, review and remove content that is violent or extreme in nature.
- eCommerce and retail companies can recommend relevant products to their customers to increase Average Order Value and overall business performance.
Today’s challenge: Finding datasets that are both expansive and accurate
While the growth of computer vision-based technologies is staggering, the truth is that companies can only do as much as their own data allows. Researchers seeking to train any machine learning models need large, diverse datasets with comprehensive, accurate metadata. They often don’t have access to this data through their own systems, and can struggle to find partners with the volume of assets they need that have the right quality, security, and legal protections.
That’s where Shutterstock comes in.
The 6 features that set our data apart:
Today, Shutterstock offers one of the largest and most accurate datasets of media content available. These high-quality, diverse datasets provide ideal training for machine learning algorithms for a range of use cases like visual search, object classification, product recommendations and image organization. Shutterstock also gives users the freedom and flexibility to create custom datasets based on their needs.
Here are 6 factors that set our data apart from the pack:
- An extensive content library of 350MM+ tagged images and videos
We provide 350MM+ images, 18MM+ videos and 1M+ 3D models with over 1M new assets added weekly to ensure content freshness.
- Diverse representation to minimize implicit bias
We source content from 1M+ contributors from 150+ countries, reflecting people of different ethnicities, gender, orientation, and ability so you can minimize implicit bias in your models.
- 7MM+ keywords in 21 languages
Our library averages 30-50 keywords per asset, and keywords are translated into 21 languages to provide flexibility and localization options.
- Accurate metadata with a robust AI and human review process
We’re the only library backed by both human and computer review to ensure relevance, with consistent metadata standards for 17+ years.
- Industry specialization in 30+ distinct categories
With our existing curated collections, we can support a wide range of industry categories like food and beverage, transportation & autonomous vehicles, animals & wildlife, clothing & apparel, travel, tourism & hospitality, etc. Or, we can create a custom dataset to meet your needs.
- Rapid deployment to get you up and running faster
Our solution enables data partners to build visual recognition algorithms faster because the underlying image data set is complete and proven to be both accurate and stable.
From broad keywords to pixel-level data: A very brief history of computer vision at Shutterstock
Shutterstock has come a long way with our computer vision capabilities. Initially, our search algorithms were powered by keywords provided by our contributors when they uploaded content. That keyword data, while useful for indexing images into categories on our site like Animals & Wildlife, Education, Food & Drink, Interiors, Science and more, wasn’t nearly as effective for surfacing the best and most relevant content needed to train AI algorithms.
For example, while indexing images in the “Transportation” category is helpful for search queries, machine learning models require a more detailed level of description (e.g. “stop signs”, “traffic lights” etc.) in order to operate successfully.
To improve upon this model, our computer vision team worked to apply machine learning techniques to reimagine and rebuild the tagging process.
Today, our technology now relies on pixel data within images. This means that at its core, our models have analyzed over 350 million images and video clips and broken them down to recognize what’s inside each image on a pixel-by-pixel basis. This includes recognizing shapes, colors, and the smallest of details — just like the human eye can.
All Shutterstock assets are then tagged with descriptive titles and 30-50 relevant keywords, providing an ideal dataset to train machine learning algorithms.
As we move forward, we are committed to actively sourcing even more diverse content to further minimize implicit bias, as well as creating new content categories and evolving our technology and human review processes to provide the highest quality, most accurate metadata available.
With Shutterstock’s datasets, your team can now easily build, train and automate these image recognition models to vastly improve your technology, and better serve the needs of your customers.
Interested in learning more about Shutterstock’s computer vision and machine learning capabilities? Visit the developer portal to create an app and contact us to get access to the complete computer vision product.
Feature photo by Roxana Bashyrova