WELCOME TO COMMUNITY EOD DATA HUB
We have a whole bunch to announce this month.
The team at Deep Analytics has been putting in a significant amount of work behind the scenes for The Hub lately. We’re almost ready to release the completely new version. The updated Hub will feature significantly streamlined automatic labeling, mobile readiness, live contests, and more. We’re currently performing the last few tweaks and tests, but it should be out in about a month.
When the new Hub launches, all current users will receive an email letting them know that the migration to the new Hub is complete. All current users will already be registered for the updated Hub. You don’t need to re-register or supply documentation again. Everything uploaded on the current Hub will also be available for labeling, reference, etc. on the new version of The Hub, so please keep uploading and tagging photos on the current Hub in the meantime. We’ll send out an email with more specifics when the time comes to make the switch.
Keep reading for information about the development process for the new mobile access version of The Hub from Deep Analytics’ Senior Software Engineer, John LaRue, new contest details, and more. As always, thank you for being part of EOD Data Hub. As Community EOD Data Hub relies primarily on crowd sourced data, the participation of Hub members is critical. Keep up the good work uploading and labeling images.
Over the past year, Senior Software Engineer John LaRue has been working on a number of Data Hub upgrades, including working with CloudFormation to make The Hub mobile friendly. Instead of manual setup, he uses infrastructure-as-code templates, improving agility and reliability. These templates allow for quick, consistent environment setup and testing. While John uses AWS CloudFormation, similar methods work with other cloud providers like Google Cloud, Microsoft Azure, and Oracle OCI, using tools like Terraform or Ansible. Any cloud setup can be scripted with APIs for easy deployment and management.
“In our application I made a template for AWS CloudFormation that defines all the components the application needs to run, using native platform services wherever possible; a database (RDS), private networks (VPC), a load balancer, web servers, serverless functions and GPU VMs. You end up with a YAML file representing the stack that you can instantiate or destroy on demand using commands like SAM deploy. When you run it, the Amazon API walks through each of the items you have defined and attempts to set it up based on your dependencies and constraints, eventually returning details for how you can reach your new endpoints.
In practice, there is a lot of trial and error before you reach the moment where your stack creates successfully. It can be quite discouraging at times, too, because the error messages can be cryptic and there are always strange restrictions and limitations about how different features can be invoked that you'll eventually bump into. Fortunately, I have not yet found a particular requirement that couldn't be worked in by some manner of template extension or code injection. It is also a useful exercise to help uncover all the hidden dependencies you've been relying on during your local development that you didn't realize would be required to deploy it in production. In this way, the resulting template goes a long way to function as a sort of living documentation for your application, dutifully outlining exactly how its components connect and what they look for at runtime,” explained John.
By templating infrastructure, John has gained confidence that once a working stack is defined, it can be consistently reproduced whenever needed. Though the process involves overcoming deployment challenges, it ultimately results in a powerful, automated system that simplifies development. Once everything is codified, John and the rest of the Data Hub team can easily experiment with features like auto-scaling, redundancy, and CI/CD integration, enabling parallel development without disruptions. Getting The Hub to work as a mobile friendly site has been a critical hurdle for the team at Deep Analytics to make The Hub as user-friendly as possible. We’re looking forward to unveiling what we’ve been working on in the coming weeks.
We’re streamlining the labeling process to save you time and increase accuracy
The team at Deep Analytics has been working on a bunch of updates to improve the functionality of The Hub and make it easier to use. One of the biggest changes in the works is to the labeling process when you upload photos. In the near future it will no longer be necessary to draw a bounding box around each target (e.g., ordnance) and select an appropriate label from the list. Instead, Deep Analytics has developed an AI model that will automatically select and label ordnance within each upload. The labels will be approved by later users to ensure everything gets labeled correctly. This will make the uploading and labeling process significantly faster, particularly for users uploading multiple images at a time. Of course, hand labeling is still an option if you prefer it, or upload something truly unique.
"Hand labeling data has always been a tedious task and at times feels like a downright waste of time. This makes crowd sourcing labels from the EOD community particularly challenging, as very few people are willing to label data for free. The Hub’s new automatic annotation pipeline aims to alleviate this problem by training an ML model to suggest new labels for images uploaded to the hub. It would be unwise to trust the ML model completely, so we've added in few new features which seek user feedback to correct it when it gets something wrong," explained Isaac Padberg, one of Deep Analytics’ engineers working on The Hub.
We talk a lot about how important crowd sourced data is for the success of The Hub, and this upgrade is a great example. Object detection models require huge amounts of data to yield accurate results. All of the photos updated to The Hub so far made up a sizable portion of the photos used to create the object detection tool. User feedback and corrections will also increase our object detection accuracy over time.
Deep Analytics has been working hard on many new and updated features for The Hub. There’s a mix of upgrades based on user feedback, and cool, completely new features. These features are summarized below.
Improved Lexicon for Both IEDs and Ordnance
We’ve gotten a lot of requests for an updated lexicon, and we think we’ve finally developed something everyone can agree on and find useful. We’ve also updated component labels and streamlined the labeling process.
Automated Labeling Powered by AI
One of our priorities with the new version of The Hub was to make it easier to use, so we’ve integrated AI to streamline the labeling process. When you upload a photo, AI will provide a suggested starting label(s) for the explosive hazards in the photo. End users can easily edit these suggested labels to provide more detail.
AI Enabled Easier User Feedback
The AI we’ve used to help label incorporates user feedback to increase its accuracy in the future. The process is quick and user-friendly. More importantly, the more you use The Hub, the more accurate the AI algorithms will become.
Community EOD Data Hub, Now on Your Device
This is a big one. You can now access The Hub from virtually anywhere using your mobile device. The new mobile site has all of the features of main site, except manual labeling, which we purposefully omitted because it would be unwieldly on a smaller screen. Now you can use The Hub to upload threats and look up information in real time, when and where you need it most.
Each month we select the most interesting recently uploaded image as a “Challenge Photo.”
Click the link below to see how many components you can label.
We’re looking for new challenge photos.
As we’ve said many times before, The Hub relies on crowdsourced data. We provide prizes to users who upload the image we use as Challenge Photos in our monthly newsletters.
To enter, all you have to do is upload images. If your image is chosen as the “Challenge Photo” in our next newsletter, we’ll send you a $20 Amazon gift card.
New Spot the AI Data Set
If you haven’t already played it, Spot the AI is part of an IWTSD project that focuses on using AI to generate realistic images of IED components, IEDs, and ordnance.
The testing is a fun and simple game. You’ll be presented with two images of ordnance, and you just choose which images are fake – could be one of the images, both, or neither. Be careful. This image collection is tricky. There are images of real and fake devices pasted onto a variety of backgrounds.
You can play the game as many times as you like. If you get an especially low score (0 is a perfect score in this case) send Morgan (Morgan.Glines@deepanalyticsllc.com) a screen shot, and Deep Analytics will send $20 Amazon cards to the lowest scoring participants when we send out our next newsletter.
Click the link below to play:
A quick labeling review.
While he Hub’s improved lexicon for labeling threats will soon be rolled out to everyone, The Hub currently uses the WTI Lexicon. Take a moment to review the WTI lexicon if you need a refresher.
How to use The Hub.
The Hub has gotten a lot of new users in the past few weeks, and to help them get started, we’re re-posting this quick video tutorial on how to upload and tag images on The Hub. We hope it will be useful to our new users, and a good refresher for anyone else.
Click the button below for our quick start guide: