Google has recently begun offering a large suite of AI tools to everyday businesses and ambitious software products that use Google platforms, perhaps from a capable Chromebook. One interesting example in the past few years is Google Cloud’s AutoML Vision. This tool to create image analysis models may soon shut down for Vertex AI’s rollout. However, you can still learn how it works and how to prepare data, which will be invaluable when using Google’s similar future services. Let’s look at the details.

What is Google Cloud’s AutoML Vision?

AutoML Vision is available via Google Cloud. It allows you to train machine learning models to classify images and recognize objects in those images. It’s part of the greater Cloud Vision platform for software developers. It includes kits to develop mobile apps and provides AI image solutions in usable formats for internal teams or consumers.

Cloud Vision is all about leveraging artificial intelligence (AI) to analyze several images at once to recognize important things like human faces, kinds of animals, buildings or cars, and so on. The service includes general AI image analysis on its own. AutoML Vision allows users to build more specific AI models for targeted tasks that can crop up in some industries. In the past, training AI models like this would take a lot of tedious human labor that few organizations can access. Google’s solution automates the labeling process to save an incredible amount of time.

Last but not least, AutoML Vision isn’t long for this world. Its time is up in early 2024, at which point users will need to start using the newer, merged AI service called Vertex AI.

Is AutoML Vision for businesses only?

It’s only for larger image-based AI projects typically needed by enterprise-level companies or ambitious start-ups. It requires a ton of cloud storage and high-tier Google business plans, among other things.

What projects can businesses attempt with AutoML Vision?

AutoMl Vision allows businesses to build AI image models via Google Cloud, so they don’t need all the hardware onsite to use complex, energy-hungry AI. But who needs to analyze huge batches of images like this outside of image generators like Midjourney? There are quite a few specialized cases where companies may want a more targeted solution, like:

• Analyzing a large product catalog and automatically labeling images: Urban Outfitters is doing this to recognize types and styes of clothing in its product catalog, something beyond more general AI.
• Tracking and identifying wildlife from an array of cameras: Conservation groups like the Zoological Society of London have training models to track wildlife species across a wide area.
• Security cam analysis to check for security incidents and perpetrators. Large-scale security firms are interested in these kinds of software solutions.
• Companies like Kewpie are training AI to monitor ingredients in their production facilities to watch for quality and consistency as supplies pass by cameras. Other manufacturers are using similar technology to watch for defects in their factories.

Plus, since organizations are tapping into Google’s image recognition capabilities (well, at least some of them), they don’t need to build out or rent expensive AI elsewhere.

How would my team prepare content for AutoML Vision?

In AutoML Vision, you must carefully prep data before using the training services. Small errors at this stage can ruin the output, which is an important part of the process. It’s also deeply embedded in Google tools, so your team must be comfortable with the common Google suite and the GCP (Google Cloud Platform) project tools.

When ready, a team creates a specific GCP project for training a model and then collects all images used for training in one batch. Images should ideally be high quality and include some outliers and “mistakes” for easier training, and companies will want a lot of them. Ideal training uses tens of thousands of images. The Zoological Society of London we mentioned above is analyzing millions of wildlife images. If you don’t have at least a few thousand images, this training may not be the best solution (despite the technical minimum of 100 images).

Teams will create CSV files to label objects and indicate the more-or-less precise coordinates of those objects in an image. This is relatively easy to put together with enough training, and Google’s AutoML services are designed to analyze CSV files as long as they’re included with the images. After the labeling data and images are provided to AutoML Vision over the cloud, teams specify the training length to train an AI model.

Additionally, teams can manually draw on images with a Google annotation tool to show what they want to be identified, which may save time in certain situations.

Can I export my AutoML Vision models?

These AI models are designed to be exported in many different circumstances, especially via Google architecture like Tensorflow Lite and Apple’s CoreML.

When is AutoML Vision shutting down, and what do I use next?

AutoML vision is currently deprecated and, as of right now, will be shutting down on January 23, 2024. That leaves no time for migration if companies haven’t started. If you’re new to AI training, begin with the successor: Google’s Vertex AI. This more comprehensive suite covers basically all of Google’s B2B AI building services.

Companies that have been working in AutoML Vision will find Vertex AI’s AutoML tools similar to use, so not much retraining is necessary. However, companies must migrate data into Vertex AI and watch for significant differences during the formatting process. Some pricing between Vision and Vertex is different. Low-end AI models, in particular, will be more expensive or won’t be available. Migration is free, but if new resources are created or new storage is required as a result, extra charges could be added on with the new pricing plans.

What does AutoML Vision cost?

It doesn’t matter. If you’re starting a new project, you’ll want to use the new Vertex AI suite instead. Let’s assume you’re still working with images. In Vertex AI, charges depend on what you do with various images. But since we’re focusing on AI AutoML prices, let’s break down what AutoML services currently cost in Vertex.

• Training for classification: $3.465 per hour
• Training for object detection: $3.465 per hour
• Training with Edge On-Device Model: $18 per hour
• Deployment and prediction services: $1.375 – $2.002 per hour
• Batch Prediction: $2.222 per hour

Advanced services, like custom-trained models for specific subjects, require specialized work and more personalized pricing.

Will Cloud Vision and AutoM Vision (or Vertex AI) take my analysis job?

It’s a common question and difficult to predict. Google would love for its AI services to be involved in everything, pinning down every bit of content for analysis. That’s not exactly feasible right now, but companies around the world are still working to understand what customized visual AI services can do for them. Claims vary from “AI will do everything for us” to “AI may have some niche solutions for us, but not at these prices.”

What we’ve seen is customized AI models in this vein taking over tedious jobs that humans aren’t good at. Watching products on an assembly line for visible defects is a great example. Not many people work jobs like that anymore. Other jobs that could see revolution include security monitoring and QA work in many industries.

Keep AutoML Vision’s successor in mind for larger AI analysis projects

Companies have only a few months to work with AutoML Vision. If you haven’t started migrating data to Vertex AI, you don’t have much time left. For newcomers, Vertex AI provides similar services in a broader package.

As more companies consider how AI could help their processes, Google is more than ready to provide time-saving, readymade solutions. The cost (in addition to the literal cost) is living entirely on Google’s platform until the AI model is ready for exporting. To learn more, check out the latest on Google’s Gemini project and how Google is using AI for Google Recorder.