Google Photos Launches

Google launched Google Photos at Google I/O on May 28, 2015, spinning it out from Google+ Photos as a standalone product with a simple and powerful offer: free unlimited photo storage for images up to 16 megapixels and videos up to 1080p, with no storage quota against the user’s Google Drive or Gmail space. Google+ had been launched in June 2011 partly as a photo-sharing platform, and over four years it had accumulated enormous numbers of photos; by separating Google Photos from the declining Google+ social network, Google preserved the photo infrastructure and its 200+ million existing users without requiring them to maintain Google+ engagement. The pricing and capacity offer was immediately competitive: Apple’s iCloud offered 5 GB free (shared across photos, mail, and backups), Dropbox offered 2 GB free, and Amazon Prime Photos (November 2014) offered unlimited photo storage to Prime subscribers but only 5 GB for videos. Free unlimited storage for full-resolution photos — at a scale Google could afford because photo storage costs had declined and because Google’s use of the photo data for machine learning purposes provided indirect commercial value — positioned Google Photos as the primary cloud photo backup destination for Android users and a compelling alternative for iOS users.

The core innovation that distinguished Google Photos from simple cloud storage was its computer vision search and organization system. Google trained deep convolutional neural networks on massive labeled datasets to recognize approximately 1,000+ object categories, scene types, and activities present in photos without any user-provided metadata. A user could search for “dog,” “birthday cake,” “hiking,” “snow,” “sunset,” “beach,” “baseball,” or “whiteboard” and receive relevant results from thousands of personal photos that had never been tagged or organized manually. Face grouping automatically clustered photos by the same person across the entire photo library using face recognition; Google required user opt-in for face grouping in most jurisdictions and did not label faces with names by default in regions with biometric data regulations (though users in the US could optionally add names). The underlying machine learning system was built on Google’s internal infrastructure: in 2014, Google had published the Inception architecture (GoogLeNet) which won the ImageNet Large Scale Visual Recognition Challenge with a 6.67 percent error rate, and the Google Photos system used descendants of these architectures trained on Google’s TPU and GPU clusters. The Assistant feature automatically created animated GIFs, photo collages, panoramas, and “rediscover this day” memory notifications from the photo library without user interaction.

Google Photos reached 1 billion monthly active users by 2020, at which point Google announced that the free unlimited storage offer would end for new photos uploaded after June 1, 2021 — photos would instead count against the standard 15 GB free Google One storage allocation. The change reflected both the commercial reality that storing petabytes of photos cost meaningful amounts even at Google’s infrastructure efficiency, and a shift in Google’s product strategy toward paid Google One subscriptions. For developers, Google Photos launched an API in 2018 (after several years without a public API, which frustrated third-party apps that wanted to access user photo libraries) that allowed reading and uploading photos, though the API scope was more restricted than the earlier Picasa Web Albums API it effectively replaced. The service’s machine learning capabilities also evolved significantly over its first five years: by 2020, Google Photos could identify specific locations from architectural context (recognizing the Eiffel Tower, the Colosseum, or Central Park from visual content alone), detect documents and make text within photos searchable via OCR, automatically straighten horizon lines, enhance image quality with machine learning-based photo editing, and provide video stabilization for shaky handheld recordings. These capabilities, all running at consumer scale on Google’s infrastructure, made Google Photos one of the most visible demonstrations of applied deep learning available to a general audience during the period when deep learning was transitioning from a research result to a production technology.