Computer Vision at Pinterest

Pinterest is an inherently visual product. In fact, it is probably the only visual product at the scale of hundreds of millions of users. The visual format of Pinterest allows the users to process a much larger volume of information, especially in visual domains as fashion, beauty, home decor. This allows for discovery and exploration and differentiates Pinterest from other platforms where users see individual answers.

Pinterest has launched best-in-class visual discovery and recommendation products that use visual input as the user query. We just launched Pinterest Lens to enable people to discover recommendations for objects they see in the real world using their camera. Last year we launched object detection, and we’re just getting started with experimentation with additional mediums like video. We currently serve 250M+ visual searches each month, and have detected over a billion visual objects in our images, but there are still many challenging problems be solved in fine grained recognition, object-to-object visual search, and large-scale visual search infrastructure.

The future for visual discovery at Pinterest is to use the unique visual dataset that is available though user curation and interaction and build the first ever Taste AI and Style AI that will understand images with their individual objects, styles and relate those to the taste of the users as derived from Pinterest interactions. Such systems will create brand new user experiences in this domain and make Pinterest a central point for visual exploration




  • Andrew Zhai
  • Dmitry Kislyuk
  • Eric Tzeng
  • Michael Feng
  • Trevor Darrell