All News | Boards | Chips | Devices | Software | Archive | About | Contact | Subscribe
Follow LinuxGizmos:
Twitter Facebook Pinterest RSS feed
*   get email updates   *

Amazon spins an Ubuntu-driven “AWS DeepLens” cam

Dec 4, 2017 — by Eric Brown 2,278 views

[Updated: Dec. 8] — Amazon unveiled a 4MP, HD-ready, $249 “DeepLens” machine learning camera with AWS hooks, that runs Ubuntu on a Cherry Trail SoC.

Amazon Web Services, Inc. (AWS) expanded its AWS cloud ecosystem with a Linux-powered deep learning camera and a FreeRTOS variant, both of which feature built-in connections to AWS and the related AWS IoT Core platforms.

The 4-megapixel, HD-ready AWS DeepLens development camera for machine learning is available for $249 pre-order, with shipments expected in April. Billed as “the world’s first video camera optimized to run machine learning models and perform inference on the device,” the WiFi-enabled camera supports a newly announced Amazon SageMaker development framework for managing the machine learning model process.

AWS DeepLens (left) and its object recognition model in action
(click images to enlarge)

From a hardware perspective, there’s nothing particularly extraordinary about Amazon’s AWS DeepLens camera. However, the “fully programmable” device has an Intel Atom x5 “Apollo Lake” SoC with Gen9 graphics, as well as 8GB of RAM, which together are sufficiently powerful to enable local deep learning processing with hooks to AWS services.


AWS DeepLens

When it ships in April, AWS DeepLens will include six sample projects that can be run as is, or can be modified and connected with other AWS services. Some of these, including face and object detection, and differentiating between a cat and a dog, are similar to those found on the TensorFlow-oriented AIY Vision Kit that Google just released for the Raspberry Pi Zero W. There’s also an activity detection project that can detect 30 different activities ranging from brushing teeth to playing a guitar. Another model lets you transfer the style from an image onto a video sequence in real-time.

Developers will be able to train models in the new Amazon SageMaker framework, and deploy them to the camera, or extend models using the AWS Lambda development environment for creating triggered actions. The camera is also optimized for connection with Amazon cloud services such as Amazon Kinesis Video Streams for streaming video to AWS and Amazon Rekognition for video analytics.

The AWS DeepLens connects securely to AWS IoT Core, as well as Amazon’s Linux-oriented, local processing spin-off, AWS Greengrass. It can also hook up to Amazon services including SQS, SNS, S3, and DynamoDB. The camera pre-installs an optimized inference engine for deep learning using Apache MXNet, and it supports other third-party deep learning frameworks. Optimized support for TensorFlow and Caffe will be available in the future.

The new Amazon SageMaker is a “fully-managed service that enables developers and data scientists to quickly and easily build, train, and deploy machine learning models at any scale,” says Amazon. Designed primarily for developers who are new to deep learning, Amazon SageMaker streamlines the process of building ML models and preparing them for training.

The “end-to-end” machine learning service also guides users to select and optimize the best algorithm and framework for any given application, and then walks them through the training process, including automatic model tuning. Amazon SageMaker can also deploy a model on an auto-scaling cluster of Amazon EC2 instances “that are spread across multiple availability zones to deliver both high performance and high availability,” says Amazon. Training models built with SageMaker can be sent to the AWS DeepLens camera with a few clicks using the AWS Management Console.

The AWS DeepLens camera runs Ubuntu 16.04 LTS on an Intel Atom, which according to this Intel blog announcement, is a quad-core Atom x5. Intel’s blog pointed to a list of Cherry Trail generation Atom x5 models, but Amazon has notified us that it is in fact an Atom x5 from the more recent Atom E3900 Apollo Lake family. Based on Amazon’s 106 GFLOPS performance claims, this would appear to be the dual-core, 1.3GHz/1.8GHz Atom x5-E3930 with 6.5W TDP rather than the 115 GFLOPs, quad-core Atom x5-E3940.

AWS DeepLens detail view
(click image to enlarge)

The AWS DeepLens is equipped with 8GB RAM, a microSD slot, and 16GB flash, which can be optionally expanded. The device has built-in dual-band WiFi, as well as dual USB 2.0 ports, a micro-HDMI port, and an audio-out jack. Other features include power and reset buttons, a power jack, and LEDs.

The device weighs 296.5 grams and measures 168 x 94 x 47mm with the detachable 4-megapixel camera mounted. The camera supports MJPEG stills, as well as 1080p video with H.264 encoding.

The camera’s baseline inference performance is 14 images/second on AlexNet, and 5 images/second on ResNet 50 for batch size of 1, claims Amazon. The device supports SSH connections, and can be programmed with Python 2.7 in addition to the aforementioned Amazon software and services.

Further information

The AWS DeepLens is available for pre-order at $249, with shipment due April 14, 2018. More information may be found at Amazon’s AWS DeepLens shopping page and product page.

(advertise here)

Print Friendly, PDF & Email

One response to “Amazon spins an Ubuntu-driven “AWS DeepLens” cam”

  1. Jesse Jenkins says:

    if the camera is detachable, what is the connector for it, ie can we attach it via a ribbon cable or some other means, as the orientation is not ideal for some situations.

Please comment here...