Google has released its in-house object detection system – known as TensorFlow – to the public, following a successful run as the search giant’s proprietary image identification algorithm.
Following a string of successful tests, and a run producing results for multiple research publications, Google has elected to make the tool available to the wider research community – and indeed, developers of all stripes – through its new, TensorFlow Object Detection API. An open source framework constructed atop the existing software, the goal is to make creating, training, and deploying object-detecting tools easier, and more accessible.
The models that come bundled with the API run the gamut; from seriously powerful inception-based neural networks to sleek, nimble models built to run on everyday hardware. To the point, a single shot MobileNets detector is bundled with the rest, and it’s optimized to run on an average smartphone, in real time without significant lag. This piggybacks off Google’s earlier announcement, introducing the world to the MobileNets line of resource-friendly vision models. These light weight wonders can handle facial recognition, object detection, and recognize landmarks, all on less-than-optimal hardware.
While modern smartphones have come light years since the flip phone era, they still lag far behind serious desktops and servers. This places developers at a crossroads.
They can run their ML models from the cloud; adding latency, and requiring a constant internet connection, which can be a deal breaker for many common applications. Alternatively, the actual models need streamlined and simplified, which leads to trade offs in utility, for the sake of an easier time deploying them.
Google has been going hard on ML models for mobile, and they’re joined by Apple and Facebook, with their CoreML and Caffe2Go frameworks. Having said that, when it comes to the public cloud, Google has a clear advantage, and it’s hardly new at delivering ML vision software, via its Cloud Vision API.
Google’s quite confident in their new software. Find out if you agree, by downloading the TensorFlow API here.