AWS DeepLens - Deep learning enabled video camera for developers

I have not had any of the setup or reliability problems others have reported. I was able to train several custom resnet50 models in sagemaker and deploy using the deeplens-object-detector as a template. Without too much of a learning curve I’ve been able to extend the logic to send recognized files to the cloud and push out notifications.The greengrass IoT framework manages deployment, upgrades, and communication to/from aws to the deeplens.If you don’t know Linux you are going to struggle. You can use several of the built in models to experiment with different aspects of ML but the real value is using sagemaker to train your on custom model. (Hint: If you do this and you use resent make sure to run the mxnet deploy script and use the model optimizer python module on the device. )The really powerful thing about Deeplens is that it’s equipped with an intel GPU that is able to do real-time object detection and take actions right on the edge of the cloud.The documentation is fairly complete but you will have to dig or read code to get the details. I find the forums helpful, there seem to be some very knowledgeable aws employees contributing.There are a number of demos available both from amazon and third parties. I agree some are better than others.This is a cost effective package for experimenting with object detection in the real world. When combined with aws ML and IoT services it is an amazing platform to build on top of. Like most platforms it does come with some complexity but it’s well worth it.My biggest complaints:1 the setup process stinks. It does not follow the model of the echo.2 any disruption of wifi/Internet connection and the device seems to hang and need power cycle to restart.

Performance is very slow for a $250 device, at times I had a lag of 1 min. Wifi kept dropping and the device would restart on its own.You can get better performance on raspberry pi for $35 after some clean up. Documentation is confusing and some of the examples don't work.May be worth trying Google's AIY vision kit ($89 on target) with deep learning models out of the box.

A few problems to install it. Documentation is not really good, a few missed instructions that one needs to discover (like the MicroSIM installation and the certifcate download). On the other hand, it is hard to find the "what next" once you deploy your first project. I discovered later the videos :/

So lets start with the basics. There are now well documented tutorials. More importantly the tutorials teach you how to connect deeplens with other aws services. That wasn't there before hence the poor reviews. You'll need to know linux, lambda and be probably some more intermediate programming concepts to make the $ spent worth it.If your just getting started with machine-learning/IOT I think this is a fantastic product because it integrates everything into one package. There is a pretty steep learning curve ~5 hrs if you going through the tutorials. But like all things aws it becomes fairly straight forward once you understand the series of tasks that need to be done each time to get things working.As far as performance goes i'm getting ~30 fps at 1080p even with compute intensive tasks. There are more then enough inputs and the screw mount on the bottom is a nice touch.Negetives: Doesn't come in blackNeed a comprehensive understanding/walkthroughDoesn't work out of the box

Because this is strictly for developers, I wasn't expecting a walk in the park. That being said, the tutorials on AWS are dated and need to be updated. There are bugs that may force you to come up with your own solutions (ex: if your SSID has a space in it, the deeplens will not connect to the wifi during the setup procedure). Because I didn't have a mini-hdmi cable, I was forced to create a guest network to connect initially and then manually configured the network afterwards (this requires familiarity with linux network configurations). This is not an end user product, so these issues are not completely unexpected.Once I set everything up, I discovered that some of the demos did not work as expected. Also, many of these demos like to send IOT messages to AWS, so prepared to be charged for usage (not very much, but still). Also, if you want to do any SSD type modelling, the GPU usage is not free on AWS so this is an extra charge as well. I sort of expected for this to be free at least for a trial period after paying the $250 price for buggy development environment.Last, but not least, all the examples steer you towards Sagemaker and other AWS cloud features, which are great, but do cost money. You don't really have to use AWS if you know what you are doing; however, if you go that route, you really don't need this product, you can just buy a Raspberry PI and a simple camera and have fun. In fact, you can probably setup 3 PIs for the price of one deeplens.In the end, I felt like this product really didn't provide me what I expected in a development environment, but your mileage may vary.

This device is excellent for anyone who wants to dive into developing either simple or advanced machine learning models. It is definitely more for educational purposes, but I found that the face detection and cat & dog recognition features are very useful for household purposes. This really is the best device on the market that developers can use to get hands on experience with machine learning without having to spend too much. I also found that the set up wasn't difficult if you watch the video posted on the DeepLens registration page. I really appreciate this device because, it is the first of its kind and it has finally allowed me to learn more about machine learning.