Google handed a new set of intelligent object detection capabilities over to the open source community as part of continued development of its TensorFlow framework.
The TensorFlow Object Detection API gives data scientists and developers access to the same tech that Google uses for its own systems, like the Nest Cam, similar items in Image Search, and street number identification in Street View. The system that Google released won Microsoft’s Common Objects in Context (COCO) object detection challenge last year, beating out 23 other teams.
Google’s first open source release includes a set of trainable models that are built for object detection. Also included are a set of parameters for those models that were trained on the COCO data set so that it’s possible for users to get started with the API without taking the time to train the models. An included Jupyter notebook will walk interested users through testing the new capabilities.
For those developers and data scientists who want to train their own models, Google has included scripts for doing so locally as well as on the company’s set of cloud services.
It’s part of Google’s push to bring AI tools to the masses, and could help accelerate the creation of more advanced intelligent systems by allowing other data scientists and developers to build on top of Google’s existing work.
These capabilities may also help give TensorFlow a leg up when it comes to mind share among people using different machine learning frameworks. Google’s isn’t the only one around: Microsoft offers its Cognitive Toolkit, while Amazon is backing Apache MXNet, and Facebook has thrown its weight behind Caffe2 and PyTorch.
One of the key benefits of this release is that Google has released an object detection system that can run on mobile devices as part of this offering. It’s based on the MobileNets image recognition models that the tech giant open sourced earlier this week.
