Google has released the TensorFlow Object Detection API, a new API that provides access to an open source framework for constructing, training and deploying object detection models. The framework is built on top of Google's TensorFlow, an open source software library that developers can use to deploy (via TensorFlow APIs) numerical computation to mobile devices, desktops, and servers.
Google is using computer vision to improve many of its products and services including Google Lens, Google Photos, and NestCam. The popularity and use of computer vision has been rising rapidly in recent years, and many companies are using deep learning to improve computer vision methods. Applications of computer vision include automatic image tagging, face detection and recognition, and visual search.
The TensorFlow Object Detection API provides access to Google's in-house object detection system. This first release of the API includes access to a selection of trainable detection models including (but not limited to) Single Shot Multibox Detector (SSD) with MobileNets, Region-Based Fully Convolutional Networks (R-FCN) with Resnet 101, and Faster RCNN with Inception Resnet v2.
The first release includes frozen weights and a Jupyter notebook for off-the-shelf inference purposes as well as local training scripts. Frozen weights are available for each of the released models, and the Jupyter notebook can be used with one of the models. Inference is a function in TensorFlow that builds the graph to the extent needed so that the tensor containing the output predictions is returned.
Developers and researchers can find more information about the TensorFlow Object Detection API on GitHub.