GooglePublish the TensorFlow object detection API, To help developers and researchers identify objects in the picture. Google focused on improving the ease of use and performance of the API, the new model released on June 16, in the benchmark test showed good performance, and has begun to apply to research work.
The object detection API contains highly dependent Inception-based convolution neural networks and simplified models that can run on simple machines. For example, the MobileNets single detector is optimized to run on a smartphone in real time.
Earlier, Google released a lightweightMobileNetsThe computer vision model family, which can be used to perform various tasks such as object detection, face recognition, and landmark recognition.
Smartphones can not handle computing resources like desktops or server-based environments, so developers have only two options. The machine learning model can run in the cloud, but this will increase latency and require the use of a network connection & mdash; in many cases this is not a good solution. Another solution is to simplify the model, so that you can casually put them on the device.
Google, Facebook and Apple have invested a lot of resources on these mobile models. Last fall, FacebookReleased the Caffe2Go framework, Used to build models that can run on smartphones, and Facebook's ownStyle TransferIs the first project to use the framework. At this year's I / O conference, Google releasedTensorFlow Lite, It is a simplified version of the machine learning framework. And at the end of the WWDC conference, Apple soonReleased CoreML, CoreML is designed to reduce the difficulty of running machine learning models on iOS devices.
Of course, because of the relationship between Google's public cloud service, it has some differentiated positioning between Facebook and Apple, and this is not Google's first release of computer vision services, such as the previousCloud Vision APIThe
The newly released TensorFlow object detection API is availableGitHubFound on. Google wants to be able to make it easy to use and achieve, so the entire tool suite are labeled, which also containsJupyter NotebookThe