Google Releases Object Detector Nets For Mobile
Google Releases Object Detector Nets For Mobile
Written by Mike James   
Wednesday, 21 June 2017

Google seems to be giving away its crown jewels. MobileNets are a small set of neural network models pre-trained to identify objects. In theory this software should be of great value, but Google is making it available for free.

A neural network is a fairly simple thing to use once it has been trained. All you have to do is multiply each of the inputs by the weights in the first layer and apply an output function to each. Repeat for each layer and when you reach the final layer you have your result. This sounds so easy and even the most feeble of processors should be able to calculate the model. However, this isn't really taking into account the huge size of the network. We may have improved our use of neural networks, but the real advance is simply in how big a network we can train - you could say networks have got bigger not smarter. Modern neural networks have millions of weights and so to compute the output of a trained net you still need a reasonably powerful processor. 

Google's MobileNet models are fully trained models that are designed to let you select a network size to suit the processor you are going to use. The idea is that they will allow neural networks to be computed on mobile devices such as phones. At the moment most mobile AI applications work by uploading images or voice clips to the cloud where more powerful processors can deal with implementing the model. Having a neural net which can be computed on the mobile device can be faster than an upload and of course it works when the mobile is disconnected. 

All of the models have been trained on the ILSVRC-2012-CLS image dataset and so they are capable of recognizing a wide range of objects. Not only do they recognise, they also place bounding boxes around the objects they have identified. 

mobilenet1

 

You can select the MobileNet you want to use according to the number of weights. Of course more weights equals a bigger network, which means more calculations but more accuracy. You can see how the different sizes of network perform in the chart below. The horizontal scale is in MACs or Multiplication and Addition operations needed:

mobilenet3

You can see that the MobileNets at their best are as good as well known models. On the other hand, I'm not sure what some of the very low accuracy nets might be used for - apps where objects only needed to be detected occasionally?

The following table gives you the performance in detail:

Model CheckpointMillion MACsMillion ParametersTop-1 AccuracyTop-5 Accuracy
MobileNet_v1_1.0_224 569 4.24 70.7 89.5
MobileNet_v1_1.0_192 418 4.24 69.3 88.9
MobileNet_v1_1.0_160 291 4.24 67.2 87.5
MobileNet_v1_1.0_128 186 4.24 64.1 85.3
MobileNet_v1_0.75_224 317 2.59 68.4 88.2
MobileNet_v1_0.75_192 233 2.59 67.4 87.3
MobileNet_v1_0.75_160 162 2.59 65.2 86.1
MobileNet_v1_0.75_128 104 2.59 61.8 83.6
MobileNet_v1_0.50_224 150 1.34 64.0 85.4
MobileNet_v1_0.50_192 110 1.34 62.1 84.0
MobileNet_v1_0.50_160 77 1.34 59.9 82.5
MobileNet_v1_0.50_128 49 1.34 56.2 79.6
MobileNet_v1_0.25_224 41 0.47 50.6 75.0
MobileNet_v1_0.25_192 34 0.47 49.0 73.6
MobileNet_v1_0.25_160 21 0.47 46.0 70.7
MobileNet_v1_0.25_128 14 0.47 41.3 66.2

Top1 accuracy is when the correct label is in the first position and top 5 accuracy is when it is in the top five labels. 

All of these models run on TensorFlow or TensorFlow Mobile. They are all open source and can be downloaded from GitHub. The team has produced a paper (listed in More Information) which explains how they reduced the number of weights while keeping the performance high.

It still begs the question of why Google is giving away such valuable software? 

Will the philanthropy end when a real killer app is on the horizon?

 

mobilenet2

More Information

MobileNets: Open-Source Models for Efficient On-Device Vision

MobileNets: Efficient Convolutional Neural Networks for Mobile Vision Applications. (PDF)

Related Articles

We May Have Lost At Go, But We Still Stack Stones Better 

The Shape Of Classification Space Is Mostly Flat

Microsoft Cognitive Toolkit Version 2.0

Why Deep Networks Are Better

Evolution Is Better Than Backprop?

TensorFlow Reaches Version 1 

 

To be informed about new articles on I Programmer, sign up for our weekly newsletter, subscribe to the RSS feed and follow us on, Twitter, FacebookGoogle+ or Linkedin.

 

Banner


Apache Hive Adds Support For Set Operations
26/07/2017

There's a new release of Apache Hive with new features including support for Set operations and a JDBC Storage Handler. 



Learning To Walk
12/07/2017

Deep Mind has been applying its deep reinforcement learning to the problem of walking - but not just any walking, this is difficult terrain. The neural network gets it wrong some of the time, but it i [ ... ]


More News

 
 

 

blog comments powered by Disqus

Last Updated ( Wednesday, 21 June 2017 )
 
 

   
RSS feed of news items only
I Programmer News
Copyright © 2017 i-programmer.info. All Rights Reserved.
Joomla! is Free Software released under the GNU/GPL License.